olivv-cs-ComfyUI-FunPack/CHANGELOG.md

Changelog

[Unreleased]

Added

• The rating's per-phrase emphasis now reaches MiniMax H3. Studio has decomposed every run into phrases, words and n-grams and scored each against the rating since 2.0, but the only thing reading that was the attn2 K/V patch — which H3 skips, having one packed self-attention stream and no cross-attention. The learning ran every generation and landed nowhere. It is now applied where H3 can act on it: log(weight) added to the attention logits at that phrase's token positions, so its share of the attention mass is multiplied by the weight. Boost only; damping a phrase you typed is opt-in. Phrases are located with the tokenizer's offset mapping rather than re-encoded through a 32B text encoder. A weighted run is a dense run — the bias is a mask, and SLA cannot carry one. Weights are relative to the strongest candidate, so strength means "how much the best phrase gets": an absolute scale put a whole prompt at x1.03, applied and doing nothing. Phrases are ranked by weight before the cap, and the prompt is located from minimax_token_tags rather than assumed to be the tail of the conditioning.

Changed

• A wired positive CONDITIONING now owns the prompt. CLIP used to win whenever both were connected, so Studio re-encoded from text and discarded what the node built — which on an i2v/r2v graph is not the same tensor, because Studio's encode never sees the reference image (nothing wires h3_references or source_image). CLIP keeps the negative, the references and phrase classification. No setting selects between them — wiring the input is the instruction, the same way a wired negative CONDITIONING has always won. "Skip Studio's positive processing" is gone; it was a switch for something the graph already said.
• Rating-driven phrase emphasis is OFF by default. It changes what the model sees on every rated run; unvalidated work that does that is something to turn on, not to discover. Overlapping spans are also bounded — their biases add, and three overlapping x1.5 spans compounded to an effective x3.4 on one token. A lone span is untouched. And how hard it pushes now comes from Studio's existing machinery rather than a new scale: the ceiling is the same 1.25 _v2_build_attn2_patch has used on LTX all along, and the ramp toward it is _v2_auto_strength, so one Awful reaches about 1.13 instead of jumping to maximum. It also measures the text that was ENCODED rather than the raw one, and resolves $variables in a phrase before matching it — phrase memory stores raw text, so a phrase could still carry $style while the encoded prompt had it resolved.

Fixed

• Learned conditioning directions survive a prompt edit on H3. The compatibility check covers the channel width and the batch, not the sequence length — which was fine on LTX, whose encoder pads to a fixed length. Qwen does not pad, so H3's conditioning changes length with every prompt edit, a direction learned at 499 positions met a conditioning of 492, and the blend threw: "conditioning memory failed ... used unblended". The whole learned direction was lost on any edited prompt. The payload is now resampled along the sequence with _resize_conditioning_sequence_like, which was written for exactly this and had no callers.
• Bounded Attention no longer replaces a wired conditioning with a text-only re-encode. It is always-on, and it works by re-encoding the prompt through CLIP and swapping the tensor for the result — which contains no reference image, because Studio was never given one. On a reference run that turned 512 positions into 197: the picture and the prompt were never both in the tensor the sampler received. Entries that arrived on a wire are now marked, and no step that rebuilds a conditioning by encoding text touches them. Every stage of _v2_finalize_conditioning is also length-checked and names itself if it changes one.
• A wired reference conditioning is no longer thrown away by the scene split. The split re-encodes every scene from text and its entries replace the output wholesale. It re-establishes H3's visual conditioning from source_image / h3_references — neither of which anything in the editor pipeline sets — so on a reference run the identity was replaced by text-only encodes and the character came out as a different person entirely. With CLIP disconnected the split returned None and the wired conditioning survived, which is why disconnecting CLIP appeared to fix it. Studio cannot split a conditioning it did not build, so it now keeps the one it was given and says so.
• A reference image's encoded rows are no longer steered. H3 tokenizes a reference INTO the conditioning — comfy marks the whole vision block as 0 in minimax_token_tags and text as 1 — and those rows are Qwen's encoding of the picture. Every steering path (learned directions, taste pull, concept deltas, the absolute store, negative_erase) lerps over the WHOLE tensor, so the picture was being moved toward a direction learned from text and the character's appearance drifted. Nothing read the tag map. The vision rows are now restored exactly as encoded, after every manipulation. Text rows still steer. Protection runs to the reference/prompt BOUNDARY rather than the image-tagged rows alone, so the <Picture n>: label between them is covered too: an r2v conditioning is laid out reference-first, and everything before the trailing run of text tags belongs to the picture.
• A reference image is no longer thrown away by Studio. With CLIP and a positive CONDITIONING both connected, CLIP won and Studio re-encoded the prompt from text. On H3 that is not a near-miss: Qwen3-VL is a causal VLM and a reference's vision block sits BEFORE the prompt, so every prompt row has already read the picture — and Studio's own encode gets no image, because nothing in the editor pipeline wires h3_references or source_image. R2V and I2V runs were handed a text-only conditioning. A wired conditioning whose minimax_token_tags contain a vision block now owns the prompt; CLIP keeps the negative and the references. Text-only graphs are unchanged.
• Clearing the editor's Negative prompt box now actually clears it. Studio falls back to a negative stored in its own settings whenever the wired one is blank, and the editor has no field for that stored copy — so emptying the visible box left an earlier negative in force with nothing on screen saying so. The editor's box is authoritative, empty included. The log now names which of the two supplied the text. The negative primitive is also written unconditionally, like the positive beside it — guarded, it kept the widget value harvested from an IMPORTED workflow, so someone else's negative encoded on every run.

Reverted

• GGUF loading is back to expanding the checkpoint at load. The quantized path through ComfyUI-GGUF loaded in seconds instead of a minute, and produced a model the sampler could not run: a packed sequence one 1344-token unit short of what the model built. Reverted whole rather than patched further.

Fixed

• The updater installs only what is MISSING, and never upgrades anything. It ran pip install -r requirements.txt, which upgrades any package below its version floor — and on an install full of compiled extensions (torch, comfy-kitchen, comfy-aimdo, onnxruntime, opencv) a numpy or transformers bump underneath them does not raise, it segfaults hours later with nothing connecting it to the update. Absent packages are now installed by name; a package that is present but older than FunPack asks for is reported with the command to upgrade it yourself, and left alone. Nothing missing means pip is never run at all.
• The stand-in tokenizer no longer downloads Gemma for a Qwen model, or reaches the network mid-run. With no CLIP wired, Studio fetched a 12B Gemma tokenizer whatever the model was — wrong vocabulary for MiniMax H3, which encodes with Qwen3-VL. There is now an H3 entry, and every source is tried from the local HuggingFace cache FIRST. Only if nothing is cached does it go online, and it says so: from_pretrained has no token, no timeout and no progress, and it runs on ComfyUI's execution thread, so an uncached fetch is an unbounded stall in the middle of a generation with nothing in the log to explain it.
• A LoRA weight that cannot fit is now dropped, not attempted. FunPack already detected and reported shape mismatches, then handed them to ComfyUI anyway — which builds the full lora_A @ lora_B delta and only then discovers it will not reshape into the weight. On a curve-form MiniMax H3 checkpoint that is a 96768x2688 tensor per block, 51 times (~25 GB in bf16, ~49 GB in fp32), allocated and thrown away while dynamic VRAM staging streams the model in. Nothing about the render changes — those adapters could never have applied.
• negative_erase could put an amplified-noise vector into the conditioning. "Keep prompt strength" restored each token's norm after the erase — but a token pointing almost exactly along the negative is left with nothing but rounding error, and restoring its norm scaled that residue by thousands. The gain is now capped: past it the token stays quiet, because it really was mostly the thing you asked to remove. A non-finite result is refused outright rather than passed on, since the sampler captures conditioning for the refinement key and would have banked the bad vector.
• Bypass now works as an A/B switch. Wire two alternatives at one input — MiniMax H3's ref-to-video and first-last-to-video both feeding the sampler's latent — and bypass the one you are not using. A bypassed slot no longer claims the input, no longer counts as a second source in guided mode, and no longer has to pass a value through for a consumer something else already feeds or that does not require one. Its own unwired inputs are also no longer demanded: it is auto-wired if that is unambiguous, and otherwise left alone. Bypassing the ONLY source of a required input still blocks, with the same message.
• Blocking messages are no longer cropped. The generation readout kept a long error on one unwrapped line, so most of it ran off the canvas. The error state now wraps, scrolls past 40vh, can be selected, and carries a copy button. The running readout is unchanged.

Added

• FUNPACK_FAULTHANDLER=1 makes a silent death talk. Installs Python's faulthandler, so a crash in native code prints every thread's stack instead of nothing — and registers SIGUSR1, so kill -USR1 <pid> dumps where a HUNG ComfyUI actually is, from another terminal, without py-spy, gdb or a restart. Off by default; costs nothing while idle.

• FUNPACK_PINNED_MEMORY caps ComfyUI's pinned host-memory budget. ComfyUI page-locks up to 90% of system RAM for weight streaming; pinned pages cannot be swapped or reclaimed, so once that budget is committed a further allocation wedges the host instead of killing the process — no traceback, nothing in the log. --disable-pinned-memory is the supported switch, but a rented image often bakes its launch command in where it cannot be edited. FunPack is imported after the budget is computed and before any model is staged, so the same control works from here: off to disable, 64 for 64 GB, 50% for half the RAM. It only ever lowers the number, and says what it did.

• The negative prompt does something at CFG 1 (experimental). MiniMax H3 always runs at CFG 1.0, so the negative branch is never evaluated and the negative prompt is dead weight. Studio ▸ "Use the negative prompt" encodes it anyway, pools it to one direction, and takes that direction out of the positive conditioning. Default mode is project — it removes only the part of each word that points at the negative and leaves the rest alone — with subtract kept for comparison. The vision span of an H3 conditioning is never touched. Off by default and unproven: expect concrete things to behave better than quality words.

• Studio ▸ "Skip Studio's positive processing". With both CLIP and a positive CONDITIONING wired, CLIP has always won and the wired conditioning was ignored. This inverts that: the wired conditioning owns the positive while CLIP keeps encoding the negative and the references. It skips shortcuts, $variables and the scene split, which is the point.

• A different sampler for the second pass. The Scene Chain Sampler takes an optional second_pass_sampler; unwired, pass 2 reuses pass 1's, which is the old behaviour. In the Editor, Second pass ▸ Own sampler reveals the full algorithm panel for pass 2 — what builds a shot well is often not what finishes it. Off, Studio mirrors the high pass's algorithm into that output, so a project that already used a second pass is unchanged. The console names the pass-2 sampler when it differs.

• Numbered reference slots. A node input can be wired to "Reference image 1" instead of to a particular file — it resolves to whatever is marked first among the image references, so re-ordering marks in the Media Bin re-points every socket without opening a node page. Numbered per kind, so marking an audio file never shifts the image slots. The picker offers one past the highest number that node already uses, and an unmarked slot leaves its socket unconnected in silence — never auto-wired to some other image, never reported.

• Export settings… now covers the sampling too — render geometry, the sampler and its schedule, the second pass when it is running, and the Studio / Chain Sampler overrides. Only the selected algorithm's settings are printed: a pass config always carries the hybrid / distilled / normalizing blocks and just one of them is live.

• Models & Pipeline ▸ Export settings… renders the loaded pipeline as a PNG: every loader with its full filename, every LoRA with its weight, the typed-in values of any custom node, and the host's PyTorch / CUDA / attention / GPU. Only inputs that were typed are printed — a wired input is marked ‹wired› rather than showing the stale widget behind the socket. The card is drawn in the theme the app is showing, watermarked with the FunPack version, and carries the same data as JSON in a tEXt chunk. The dialog previews it with Download and Copy image.

• Quality sharpness now works with a stock KSampler. The unsharp mask that recovers fine detail lived inside Hybrid Euler 2S and Distilled Flow only. It reads the current x0 prediction, the previous one and the step's sigma — nothing that needs a sampler's loop — so it now runs through a denoiser proxy on any sampler_name. sharpen last % sets the window as a fraction of the schedule, the same meaning as Hybrid's high quality pct. Audio on a packed AV latent is excluded, as in-loop. Off by default; a sampler that cannot be wrapped keeps sampling and says so.

• Settings ▸ Custom Nodes: install, update and remove ComfyUI node packs. ＋ Add node asks for a repository URL and clones it into custom_nodes, installing its requirements.txt if it has one; each row offers Update (a fast-forward pull, refused if the pack has local changes) and Remove. It is three git operations, not a catalogue — you supply the URL, and nothing about the repository is vetted first, which the dialog says. Removal names the full path it is about to delete and cannot reach outside custom_nodes: the name must be a single path segment that resolves to a direct child directory, so a symlink cannot redirect it, and FunPack cannot delete itself. Node packs register at import, so the panel says a restart is needed rather than restarting under you. Check for updates fetches each pack's origin and shows how far behind it is, on the row and on its Update button. It is a button rather than part of the listing because it costs a network round trip per pack; the fetches run four at a time, and a pack that cannot be compared says why (detached HEAD, no remote, origin unreachable) instead of quietly reading as up to date.

• FunPack's loaders take .gguf files, for diffusion models and text encoders alike, and a text-encoder list may mix a .gguf with .safetensors slots (the usual LTX-2.3 shape). Core's extension set has no .gguf, so those files were on disk and invisible to every picker. gguf is now one of FunPack's requirements, so a fresh install reads GGUF containers with nothing extra to set up. The runtime is whichever backend is present: ComfyUI-GGUF keeps the weights quantized in VRAM, which is the point of GGUF; failing that, the gguf package dequantizes at load — the file loads, but at full size, and the status output says which of the two happened. With neither, the loader names both remedies instead of failing obscurely. Model-family detection reads GGUF containers too, using the same architecture signatures. A GGUF renamed to .safetensors is recognised by its container magic and loaded correctly rather than failing in the safetensors parser with a UTF-8 decode error. When ComfyUI-GGUF refuses an architecture it has no handling for (MiniMax H3, today), the load falls back to dequantizing rather than stopping, and says both what refused and what happened instead.

• Simple mode's Advanced settings button opens the panel it names. It slides in the Editor's own Properties column — which does nothing at all when that column was collapsed in Editor mode, since Simple mode hides the dock tabs that would uncollapse it. Opening the panel now uncollapses it, and the saved dock state is handed back on the way out of Simple mode. The panel also shows its folded sections again: they are hidden to keep the pinned column tidy, which left a button called "Advanced settings" opening onto the few rows that are not in a fold.

• Pin up to three shortcuts to the timeline toolbar. Settings is a window with a sidebar, and its useful places — a specific node's page, one Engine category — are several clicks deep: fine to walk once, tedious while dialling something in. 📌 Pin to a button in the Settings header puts whatever is open into one of three slots, chosen from a dialog that shows what each slot currently holds and asks before replacing one. Pinning while a node is open pins the NODE, and pinning inside Engine pins the CATEGORY — "Sampler algorithm", not "Engine" — since the section itself was never the slow part. Slot 1 is always the leftmost button and slot 3 the one nearest Assets; an empty slot closes up and the remaining buttons keep their own numbers, so a shortcut's position never moves. Hovering says where it leads. Pins ride with the project, like the other editor preferences. A node opened this way is a destination, not a stop inside Settings: Save and Cancel both close the window and return you to the editor, rather than leaving you in the Models list you never asked for.

• Reverse a clip, audio included, in + Add → Effects. Unlike the other clip effects there is no preview-side equivalent, so the monitor switches that clip to a server-rendered segment — what you watch is the reversed encode, not a forward one standing in for it. ffmpeg holds every frame of a reversal in memory, so clips past a frame limit are refused with the count, the limit, and the fix, rather than being rendered forwards silently.

• Clip geometry on the timeline: flip, crop, and fill-frame, in + Add → Effects alongside the existing zoom/blur/fades. Flip horizontal and Flip vertical mirror the clip; Crop edges trims a percentage off each side and rescales (a punch-in that composes with Ken Burns rather than replacing it); Fill frame covers the output frame and crops the overflow instead of letterboxing. The preview and the render share one filter definition, so what plays is what renders. Flips and fill are switches — applying one again turns it off — and the timeline clip now shows a chip listing the effects on it, which clears them all when clicked. Remove all effects is also a preset, since Ken Burns previously had no way off. Presets are back-filled into libraries created before they existed.

• SLA block-sparse attention for MiniMax H3, as a value in the FunPack Diffusion Model Loader's attention list rather than a node to wire. ComfyUI ships no sparse-attention backend for H3, which is why lightx2v's SLA turbo LoRA gives no speedup on its own — the LoRA is the adaptation to sparsity, not the acceleration. Roughly 3.7x the attention throughput at 768p/15s. It is a toggle beside the attention backend, not one of its values: SLA takes H3's long packed self-attention and the chosen backend (sage3, int8, flash) takes the text refiner, masked calls and any trailing dense steps, so the two compose. Five settings folded under Advanced and validated at their defaults; skipped with a stated reason on anything that is not H3 or without Triton, leaving the chosen backend installed on its own. Kernel and block map vendored from LightX2V (Apache-2.0) via ComfyUI-H3-SLA-Attention (MIT).

• MiniMax H3's latent upscaler loads without its custom node pack, so second_pass_op and segmented detailing work on H3 — the operation needs an upsampler whose latents are the model's width, and the only published 24-channel one shipped behind its own node. Architecture and normalisation statistics are read off the checkpoint; verified bit-identical against the reference implementation.

• H3 scenes continue from the previous one. The previous scene's last latent frame becomes the next scene's frame-0 keyframe pin — the only continuity conditioning H3 accepts. A carried latent tail is not conditioning on this model, so the seam matched while the rest of the shot knew nothing about the scene before it.

• A resample factor for the between-pass operation (second_pass_upscale, 1.0-4.0). Only upsamplers that take a factor honour it: MiniMax H3's resizer does, Lightricks' LTX one is a fixed 2x network and reports that it ignored the value rather than rounding quietly. Latent width and height snap to even, since a patchified model cannot take odd.

Changed

• FunPack's own nodes now say what they are. A card in the Models shelf reads 🧠 Diffusion model / 🔤 CLIP model / 🎞️ VAE / 🧩 LoRA / 🔑 Refinement key, tinted to read as built-in, instead of showing the class name. Third-party nodes keep their class name — the difference is the signal: these are the pieces that come wired.

• The model family is detected from the checkpoint instead of chosen. Selecting LTX while loading an H3 file wired the entire graph for the wrong model, and the mismatch surfaced as a stray port rather than as a family error. The family now comes from the diffusion model's own safetensors header — key-name signatures only, the same ones ComfyUI's model_detection uses, so it costs the same on a 40 GB file as on a small one and never loads weights. Changing the checkpoint rewires the pipeline and migrates the project's frame geometry, which previously only happened if you went through the wizard.

  A file that cannot be identified proposes no family: the previous one stands and the panel says why. Nothing silently becomes LTX. A video-only Lightricks checkpoint is wired on the AV graph, as before, but is now named as video-only so an empty audio branch is expected rather than mysterious.

• Settings the loaded model cannot use are no longer shown. MiniMax H3 drops Bounded Attention, Best-FaceID (with its four sub-settings) and v2a_grad_scale; LTX drops h3_audio_clock. Previously each was offered, left switchable, and then reported once per generation as inert — a control that is not offered needs no explanation. Stored values are untouched, so switching family back restores the setting. The main-window chips for these are gone with them: "turn it off in Settings → Engine" named a row that no longer renders.

• Operations that fail now say so, and standing conditions stop repeating. Two rules that fight each other unless they share one mechanism, so they now do (funpack_log.py). A failure reports what was attempted, why it stopped, and what the output looks like as a result — collapsed to one line per run, because these fire inside per-step wrappers. A standing condition ("this model is not LTX", "H3 has no cross-attention to hook") is stated once and restated only when it stops being true, instead of on every generation.

  Newly audible: anchor pin restore, guide and mid-scene-guide append, velocity bias, quality sharpness, the audio-protection mask, the audio clock correction, momentum guidance, template resolution match, learned-direction steering, conditioning memory and repel, Absolute steer, and the temporal-style classifier. Each of those could previously fail and leave a run that looked exactly like one where the feature was switched off.

Fixed

• Updating FunPack in the app installs the dependencies the update needs. It pulled new code and never new requirements, so a release that added one left the node pack unable to import with nothing said. When an update touches requirements.txt, pip runs against the interpreter ComfyUI is using — never a bare pip, which in a venv installs somewhere else entirely — before the restart, and a failure names the command to run by hand rather than turning a completed update into an error. Update and branch switch also moved off the event loop; a fetch over a tunnel plus an install was freezing every stream in the meantime.

• A LoRA that matches by name but not by shape now says so. Key matching proves a LoRA is for a model, not that its weights fit it; a mismatched pair is dropped during the merge, which was visible only as one generic warning per key while FunPack's own status line still read like a clean load. The count now reaches that status line. The common case is named outright: MiniMax H3's pruned "curve-form" checkpoints read adaLN from a compact time-curve basis, so a turbo LoRA trained against the full-width model matches all 51 adaLN keys and merges into none of them. They cannot be projected — the basis that built the table is not in the checkpoint — so the message says that and points at a converted LoRA.

• Context windowing and ALG's guide blur are switched off and hidden on MiniMax H3. Core's windowing unpacks the LTXAV stream specifically and measures its window on LTX's 8x latent ratio; ALG's guide blur acts on guide frames appended to the latent, and an H3 guide is a condition row, so that tail is always empty. Context windowing also reported the wrong reason on H3 — it blamed the ComfyUI version for what is a model difference.

• JoyAI-Echo is switched off and hidden on MiniMax H3, where it was doing damage. The memory bank places frame i at sequence position i, but H3's packed layout pins only the first or last frame — so every frame past the first was refused, and the one that landed replaced the scene's i2v anchor, because H3 keys pins by frame index. Not an inert toggle: a harmful one. The sampler now forces it off on H3 and says why, and Engine Settings stops offering the whole group.

• JoyAI-Echo's tooltips now say it requires the JoyAI-Echo LoRA. Without it the injected memory frames change nothing on any model, because the base weights were never trained to read them as memory — the controls read as a working feature.

• Closing Models & Pipeline mid-edit no longer leaves its edit buffer behind. Escape, the ✕, or switching to another section tore the section down without ending the edit, so the next visit inherited a dirty flag and a baseline belonging to a config that had already been replaced — a spurious "unsaved changes" prompt, and a Cancel that could splice stale slots into freshly loaded ones.

• Prompt templates can be renamed, deleted, and turned off. The Composer's Templates bar applied a template and immediately forgot it, so there was no state to leave and no way to leave it — and clearing the prompt box by hand did nothing, because an empty global prompt is refused (an empty parse would wipe the timeline on a stray keystroke). The bar now shows which template is applied, offers — None — to clear the whole global prompt — anchor, transitions and scene texts — as a deliberate, undoable action, and gives the applied template Rename and Delete. Saving pre-fills the applied name, so updating a template is the default gesture rather than a trick. Deleting one leaves the prompt alone.

• "No, I'll use my own pipeline" left the setup modal on screen. Picking a model family starts a chain of requests that reopens the modal with fresh prerequisites; dismissing it mid-flight closed it, then the reply landed and put it straight back — and since opting out installs nothing, the prerequisites were still missing, so it always reopened. Closes the user asks for now supersede anything in flight. The choice is also recorded before the save round-trip rather than after it, and a save that fails says so instead of leaving the project quietly still on the built-in pipeline.

• An output can be wired to several destinations while adding a node, not only after. The Add-node panel offered one destination per output and told you to finish the job on the node's page; it now uses the same multi-destination editor the node page does, with the same rule about what the built-in pipeline will honour.

• Double-click an image in the Media bin to make it the selected clip's anchor. Dragging was the only way in, and it could not be relied on: an <img> is natively draggable, so inside the card the browser started its own image drag whose payload every drop target rejects — worse the larger the image. The image now opts out of dragging, and the drag ghost is the thumbnail rather than a snapshot of the full-size card.

• Preview playback could stall at "Video is loading…" and stay there, with the transport buttons dead. A media reset leaves a <video> with no metadata and no load running, and every recovery path keyed on the element having errored — which a reset does not do. Play went down the same wait-for-metadata branch, so pressing it changed nothing. Elements are now checked after a reset and restarted if no load followed; Play restarts a stalled one outright.

• A linked input driven by Project · Prompt (global) claimed to encode "the same text Studio would" and did not. Studio appends the postfix to every scene itself, so the global prompt is strictly less than what it encodes — and the postfix is usually where audio and style directions live, which went missing with no clue. The hint is now accurate per source, and points at Project · Prompt + postfix when a postfix is set.

• Studio said nothing when both CLIP and a positive CONDITIONING were wired to it. CLIP wins and the wired conditioning is never read — but the ownership label that knew this was computed and discarded, so a graph feeding Studio from an i2v node looked like it was working while the prompt path quietly supplied everything. Now stated once per run and carried in the encode status.

• The late-step gate every rating-driven mechanism shares was measuring the wrong thing on H3, so embed guidance, score slider, DynaShift and output guidance were inert or nearly inert while reporting themselves active. max(0, 1 - 2*sigma) reads sigma as schedule progress, which holds on LTX and fails on H3: its schedules are shift*t / (1 + (shift-1)*t), so a large shift keeps sigma high until the final leap. Measured coverage of the old gate — shift 6 / 4 steps (turbo): 0 of 4; shift 12 / 12 (H3's default): 0 of 12; shift 3 / 20: 4 of 20, the first two at gate 0.14 and 0.31. The gate now reads position on the schedule's own base grid, recovered from the schedule rather than from a shift constant (the shift is only reliable when MiniMaxH3SigmaShift is wired, which video sampling does not require). LTX keeps the gate it was validated with. Every run now reports its steering window, and says so outright when nothing will steer.

• Embed guidance crashed every H3 run that reached a steering step, and score slider silently did nothing on the same models. H3 refines the conditioning inside extra_conds, so the DiT consumes 5376-dim hidden state while the taste store captured the raw 5120-dim text conditioning; embed guidance raised the size mismatch, score slider caught it and returned the base prediction every step. Learned directions are now carried into the consumed space through the model's own text preprocessor, once per scene. DynaShift's prompt-similarity weighting was affected too — it compared the raw banked prompt against the refined one and weighted every negative equally. The value function had the same fault one layer down: fitted on raw conditioning, it was handed the refined tensor and threw every step, so the run steered on the fixed direction while the report claimed the value function was driving. It is now asked for its gradient in the space it was fitted in, once per scene rather than once per step.

• A resolution-changing second pass no longer fails on MiniMax H3. A keyframe pin is packed as condition rows, so its token count belongs to the grid it was encoded on, and upscale_2x handed pass 2 a pin from pass 1 — value tensor of shape [168, 96] cannot be broadcast to indexing result of shape [672, 96]. The pins are resampled onto the new grid rather than dropped: on H3 the pin IS the anchor, and pass 2 has to keep holding it.

• Multi-scene chains work with a second pass again. second_pass_op="upscale_2x" hands back a scene at twice the latent size, but every later scene is still built from the latent template at the original size — so the carried overlap frames, the anchor's continuation, the soft join, the JoyAI memory frame and per-scene guide sources were all being spliced into a chunk on a different grid, and the second scene died on the shape mismatch. Anything crossing a scene boundary is now brought back to the template's grid on the way. Each scene still samples and outputs at 2x; only the carried material is resampled, and only downwards — the direction that survives it, since those frames exist to say "continue from here" rather than to carry detail. Runs with no resolution-changing op are bit-identical.

• The About panel reads FunPack 3 "Auspicious Asparagus" rather than putting the major on the codename line.

[3.5.1] "Auspicious Asparagus" - 2026-08-11

Compatibility and polish on top of 3.5.0. LTX-2.5 works, and the two places it would have broken silently are fixed. Releases now carry a codename per major version.

Added

• Release codenames, shown in Settings ▸ About.
• About reports the machine ComfyUI runs on: chip, memory, GPU (name, VRAM, compute capability), free disk, OS, ComfyUI, Python, torch, CUDA, and which fast-attention backend is installed. On a rental that is the host, not the browser.
• A schedule for every sampler. Steps and scheduler belong to the pass, not to the KSampler branch, so any sampler can now be given a computed schedule instead of only a hand-typed sigma list. The frames field became usable in the same pass.
• ALG on any sampler. The i2v anchor blur used to be locked to FunPack's Distilled Flow sampler. It now runs on whatever sampler is wired — a stock KSampler, Hybrid Euler 2S, a two-evals-per-step sampler like heun — by lifting the guidance out of the sampler loop and onto a denoiser proxy driven by the step's sigma.
• H3 audio clock. MiniMax H3 denoises video and audio on two different flow schedules but hands the sampler one sigma grid, so the DiT reconciles them with a start-of-step slope that badly overshoots on a few-step schedule — heard as distortion. h3_audio_clock swaps that for the chord actually spanning the step. One scalar multiply, no extra model call. It works with stock ComfyUI samplers too, and stands itself down with a console note on samplers that evaluate twice per step rather than guessing.
• Projects remember your Editor settings, so a fresh rental does not reset your preferences and shortcut revolver.
• Every Engine setting says what it does, what it costs and what it needs, in one line.

Fixed

• LTX-2.5 compatibility. 2.5 reuses the same model classes behind new config flags, so nearly everything binds unchanged — but two places assumed 2.3 specifics and would have failed silently. The video/audio conditioning split was a hardcoded width table; from 2.5 on ComfyUI reads those widths off the checkpoint, and an unrecognised pair made the split return the full width, steering the audio text context along with the video. The only symptom would have been degraded audio. The split is now measured off the live model, with the table kept as a fallback. Separately, decode-time noise reached only the conv decoder: 2.5's diffusion decoder takes no timestep and seeds itself, but is an nn.Module, so the settings were accepted and then ignored — a knob that read as live and did nothing. It is now capability-tested and says so when it cannot apply.
• cut_opening_frames no longer leaves a noisy first frame. The cut was made in latent space, and the LTX video VAE is causal: latent frame 0 is the temporal origin, while every later frame was generated as a continuation. Slicing the front off the latent promoted a continuation frame to position 0, which then decoded with origin handling it was never generated for. The cut now happens on decoded pixels, after a decode that saw every frame in the context it was sampled in — the way MiniMax H3 already did it. The count was wrong too: a latent cut could only remove whole latent frames and shortened the clip by a different amount than the span it removed. It is exact now, N means N. Video comes from the IMAGES output on a cut run (the Editor already wires it that way); the latent's audio stream is cropped by the same amount of time so sound and picture still start together.
• Guide keyframes are dropped when second_pass_op changes the resolution. They are recorded as token indices into pass 1's grid, so upscale_2x left them addressing the wrong tokens — which recent ComfyUI rejects outright and older builds mis-placed silently. Pass 1 still uses every guide; only pass 2 loses them, and the scene report says so.
• One ALG control instead of two. The Distilled Flow panel had its own switch while the chain sampler's anchor blur did the same thing on any sampler and already drove it — two controls for one behaviour, with precedence depending on which you had touched. The Editor now shows one. Projects carrying the old switch are migrated with their strength and threshold, and told so.
• H3 image conditioning actually reaches the model. Two independent breaks, both silent, both ending with a generation that ignored the input image entirely and left the prompt to carry the whole shot. The Movie Editor always splits a run into per-scene conditionings, and that split re-encoded every scene without the anchor image or the ref2va references — so neither the keyframe pin nor the <Picture i> presentation ever reached the sampler, even for a single-scene project. The split now carries the anchor onto the opening scene and the references onto every scene. LTX projects are untouched: there the anchor travels in the latent, and adding vision to scene 1 would change existing output.
• A wired MiniMax H3 Image to Video node no longer throws its image away. That node carries its first/last frame pins on its CONDITIONING output, which this pipeline dropped because the sampler's positive comes from Studio. New optional Chain Sampler · h3_keyframes input (auto-wired for H3 projects) salvages the pins — first frame onto the opening scene, last frame re-indexed onto the closing scene's own final frame.
• Spare H3 token tags are trimmed, not thrown away. An image prompt comes back one tag long, and treating the surplus as corruption discarded the whole vector — which on exactly those prompts left the DiT modulating the picture as if it were text.
• Projects remember which model family they are for. Every project on disk had no family recorded while the global default said MiniMax H3, so the builder read them as LTXAV and routed the audio VAE at a node H3's graph never builds. That is the reported "phantom Audio VAE port": one root cause, four separate bugs, all fixed.
• A wired positive_conditioning satisfies Studio's CLIP requirement, and an fl2va/ref2va node satisfies H3's AV latent requirement — neither should have been blocking generation.
• Warn only when Studio has no conditioning source at all, rather than whenever one particular input is empty.

Changed

• Guide strengths span the full 0..1 range. Mid-scene guide, identity pin, prior-scene guide, the per-guide strength field, and the JoyAI memory floor were clamped to 0.25–0.5. That band is the measured audio-safe sweet spot, not a physical limit; it stays in the tooltips and out of the code.
• Project texts reach nodes that encode on their own, expanded — so a custom encoder in the graph sees the same prompt the pipeline does.
• The H3 v2a_grad_scale warning chip is gone. The knob does nothing without JoyAI audio memory on any model, so flagging it as an H3 limitation blamed the wrong thing.

[3.5.0] - 2026-08-04

A second model family. FunPack was built around LTX-2 / LTXAV and now supports MiniMax H3 as a first-class alternative: a project picks a family, and everything downstream — which nodes the graph emits, which model files the setup asks for, which sampler settings are even applicable, what a valid scene length is — follows from that choice. Nothing about LTX-2 changes.

Added

• MiniMax H3 is live. H3 support was written against ComfyUI's pull request; the model has since merged (ComfyUI v0.30.0) and the weights are published, so the pipeline setup now offers it as a real, released family with the actual filenames to download. Nothing about LTX-2 / LTXAV changes — a project picks a family, and both are first-class.
• i2v anchors work on H3. LTX anchors an image by writing it into the starting latent; H3 has no such path — its anchor is a condition row packed beside the text. A scene's anchor image (and the Cutting Room's per-scene anchors) now becomes H3's own frame-0 keyframe pin, so image-to-video actually conditions instead of quietly generating from noise. A first-frame anchor and a last-frame guide now coexist rather than overwriting each other, which is exactly H3's first-and-last-frame mode.
• Reference images can be encoded at full detail (Settings ▸ Engine ▸ References): Match output size as before, or Max detail (2048px), the reference pipeline's own sizing. Reference rows ride through every sampling step, so max detail is not free.
• Frame geometry follows the model. Frames per scene snap to the chosen family's grid (LTX 8k+1, H3 17k+5) in both frontends and in the graph builder, and H3 renders at its fixed 24 fps. Picking a model family brings the project onto its geometry and says so.
• The model family is the first question in Pipeline setup, not something to find in Models afterwards — it decides which nodes the graph emits and which files to download, so everything else follows from it. Reachable later via Models ▸ Model family ▸ Setup…
• Reference media is marked in the Media Bin and wired like any other input. Any image, video or audio takes an R mark beside the identity pin, mark order is the numbering the prompt uses, and each marked item appears as an input source in Models & Pipeline for every socket its kind can fill. At generation the files are copied into ComfyUI's input folder and one loader per reference is injected, chosen by what the destination socket asks for. Replaces the Reference media tab: references are wired now, not configured.
• Reference video clips work, decoded once at 24 fps into both views the model needs — the 2 fps timestamped frames the text encoder sees and the full-rate latent block the DiT packs — with the clip's own soundtrack riding along as its <Audio j> label. Capped at 15s, because reference tokens ride every sampling step.
• The KSampler pass can build its own schedule from steps + a scheduler, the way ComfyUI's own scheduler does. A SAMPLER object is only the step function, so this pass previously had no schedule at all unless you typed a sigma list by hand. The schedule dropdown is the switch between the two: use_user_sigmas (the default) runs the Sigmas field as typed, anything else computes — so a hand-written schedule can stay parked in the field and be switched back to without retyping it.
• A live elapsed timer on the Generate buttons — "▶ Generating (44s)" — counting from the press, in both frontends. After a reload it re-attaches to the running generation and says it is counting from the reconnect rather than claiming a total it never saw.
• A non-finite latent now stops the run where it happened. A chunk that samples to completion and returns NaN/Inf used to sail through everything and surface as an ffmpeg AAC error naming nothing that produced it, after the whole montage had been paid for — and the video half never complained at all. The sampler now checks both the incoming latent and the returned one and says which it was, with the scene label, the stream, how many values, and the causes worth testing in order. The hand-typed sigma schedule is checked too: an interior zero or a repeated value is an instant Inf, since the solvers divide by sigma.
• A node's own prompt field can be driven by the project prompt — the linked-input "Driven by" list gained Prompt / Negative prompt / Seed, filtered to the widget kinds each can legitimately fill.
• Warnings arrive before a generation is spent, not after. Beside Generate: when Best-FaceID identity transfer cannot do anything, when a scene's source mode wants an anchor image but none is picked, and when a sampler setting is inert on the chosen model.

Changed

• The sampler-name list comes from ComfyUI instead of a hardcoded eight written before the rest existed — that had been hiding 36 samplers, including res_multistep, which is what ComfyUI's own MiniMax H3 templates use. A saved name missing from the live list stays selectable rather than being swapped for the first option.

Fixed

• The log panel no longer eats your selection. It rewrote its whole body every 1.5s, which drops whatever you had highlighted — copying three lines out of a running log was a race against the next tick. Updates are now withheld while a selection is held, an idle log doesn't touch the DOM at all, and there is a ⏸ Pause toggle plus copy-just-the- selection.
• A core override left behind by a model-family switch is no longer emitted. Overrides are saved per core id, not per node class, so switching family left the replaced node's input names on its successor — invisible in the Models panel, fatal at generation (VAEDecodeAudio.execute() got an unexpected keyword argument 'audio_vae'). The builder refuses to write an input the class does not declare, and the family switch prunes them.
• H3 token tags are reconciled with the conditioning Studio hands on. The token count moves with both the prompt and the resolution, and a mismatch was a bare "list index out of range" at step 0 — which is why it looked intermittent and why editing either one made it come and go.
• H3 decodes audio from the sampler latent, matching ComfyUI's own reference graphs rather than handing the audio VAE a different object than they do.
• Autogrow list entries are addressed by their dotted socket id (ref_images.ref_image_0). The bare name is rejected by validation, or slips through and arrives at execute() as an unexpected keyword. Configs saved under the old name are renamed in place on load.
• A saved model file that isn't on this machine no longer blocks the whole prompt. An ArcFace projector name with models/loras empty made ComfyUI reject everything, even with Best-FaceID switched off; editor-supplied values now fall back to the node's default with a visible note.
• The i2v guide stack works again — a lost dataclass had left it dead.
• The between-pass sharpen no longer smears through a box filter on the way back down.
• The in-app updater survives an upstream history rewrite. git pull --ff-only cannot cross one, so the update dead-ended in a git hint; it now asks whether any local commit's content is actually missing upstream, realigns when nothing can be lost, and refuses with a count when there is real local work.
• A video reference lost its soundtrack between Studio and the sampler. Studio had already announced the track to the text encoder as <Audio 1>, but the reference list it handed on dropped the field — so no audio rows were packed for it and every later <Audio j> in the prompt pointed one reference earlier than written.
• Easy Gen never knew it was on H3. Its model config lives on the project rather than in a separate slot, which the shared capability check did not read — so every H3-specific warning silently reported LTX and never appeared.
• An off-grid scene length on H3 failed with arithmetic instead of an explanation. The length is now snapped to the model's grid, and a genuine template mismatch names the latent node and the number it needs.
• The run says which H3 checkpoint its conditioning needs. H3 ships as two DiTs — fl2va (anchors/keyframes) and ref2va (reference media) — that load identically and never reject each other's conditioning, so a mismatch only shows up as a poor generation.

Maintenance

• The 22 chronic test failures were stub collisions, not redundant tests. Every one was a real test of a live feature: 18 failed only in a full-suite run because each module built its own partial comfy stub and the first to import won. The package tree is now built once and shared. The suite is green and every file passes alone and in reverse order.
• Dead code removed: the orphaned wildcard engine, refiner-state template cluster, unused template payload/summary helpers, and unreferenced movie_editor backend helpers.

[3.4.1] - 2026-07-31

Added

• A second sampling pass on the Chain Sampler. Give it a schedule and every scene is sampled twice: pass 1 runs the main sigmas in full, then pass 2 runs its own schedule in full, starting from pass 1's finished clip. Nothing is cut short and nothing is derived — total steps are simply the two added up — and pass 2 re-enters through ComfyUI's ordinary img2img noise scaling, so its first sigma is a literal strength dial (0.8 reworks the shot, 0.4 polishes, 0.2 is detail work). The i2v anchor stays pinned across the split. In the Cutting Room it is one field, Second pass schedule, under FunPack Studio ▸ Sampler algorithm: typing a schedule enables the pass, clearing it turns the pass off.

• An optional operation between the two passes (none by default — nothing runs unless you pick one). sharpen runs one forward of Lightricks' trained 2× latent upsampler and resamples straight back to the original size: no video-model calls, a fraction of a step, and pass 2 then re-denoises the result, which is what makes it stick. upscale_2x keeps the 2×, so pass 2 runs at four times the pixels and the scene decodes at double resolution (3–5× the cost of the second half, and the i2v pin is dropped — the scene report says so). Both use the same upsampler file segmented detailing does, and the model picker now appears alongside the operation instead of only under segmented detailing.

• Cut the opening off i2v scenes (cut_opening_frames, Chain Sampler ▸ Timing & Seed). The anchor is a pinned frame at position 0: it carries identity, style and composition better than anything that weakens it on the way in — and it is also literally the first frame you see, so every i2v scene opens on the exact reference still. The scene is generated exactly as normal, anchor at full strength, no extra sampling, and the opening is then cut off the finished clip: an i2v generation that reads as t2v. Nothing is regrown, so the scene comes out that much shorter and the audio is cropped to match.

• Context windows for scenes longer than the model's comfortable window (ComfyUI core's own mechanism, audio-aware on LTX — nothing ported). Engages only past the window length, so shorter scenes pay nothing.

• The progress readout says what is running, not just a step count: sampling 17/26 · scene 2/3 · pass 2 of 2. The second pass also announces itself on the ComfyUI console the moment it starts, rather than only in the run report afterwards.

• Prompt $name variables in Easy Gen — the shorthand-for-a-full-phrase layer the Cutting Room's Composer already had ($vid = "High quality, high fidelity realistic video, motion blur, cinematic fog"). A $ Variables button in the prompt bar opens its own window with editable name/value rows, a count badge, and a live warning for names used but never declared or for a variable that references itself. Same project.variables field as the Cutting Room, so a project carries its variables between the two UIs; substitution still happens inside Studio dead-last — after shortcut expansion and after the scene split — which is what makes a $var work inside a shortcut's replacement while a value containing a comma or a trigger word can never move a scene cut.

• An Interrupt button in Easy Gen. The progress panel now carries the same ■ Interrupt control the Cutting Room's player has — Easy Gen previously had no way to stop a run short of the ComfyUI queue. A stopped run reports "Generation stopped." instead of a failure (an interrupted job is recorded as an error with no media, which is indistinguishable from a real crash at the API level), and any partial media it did write is still shown.

• Easy Gen re-attaches to a running generation after a page reload — the Cutting Room already did.

Changed

• Variable values are now auto-growing text boxes (both UIs) instead of one-line inputs — a long phrase was effectively uneditable in a single-line field.
• Plateau step-cache and context windows are mutually exclusive — the cache cannot tell one window from another within a step, so it is skipped with a note in the scene report.

Fixed

• Context windows never worked. The schedule names were ComfyUI core's spelled backwards (uniform_standard where core says standard_uniform), so every choice except batched raised a ValueError out of the sampler and failed the whole render. Core's names are the choices now; the old spellings are still accepted and mapped onto them, so a saved project keeps generating. An unknown name is refused with a reason naming what your ComfyUI accepts, instead of escaping as an exception.
• The same call also passed a keyword core's window handler does not always take, which was reported as "ComfyUI core too old" — silently switching off a feature that core fully supports. Unsupported keywords are dropped now, with a note naming the one setting that degrades.
• A rejected prompt now says why. ComfyUI's top-level error for a validation failure is a fixed blob ("Prompt outputs failed validation", no details); the part naming the node, the widget and the bad value is in node_errors, which the editor never read — so every rejected prompt reported the same nothing. Both are shown.
• Bypass no longer refuses a node over an output nothing uses. A node that emits an extra output the graph never wires (an IC-LoRA loader's latent_downscale_factor, say) could not be bypassed at all, because a passthrough was demanded for every output rather than for the ones actually consumed.
• The progress bar counted only the first pass, so a second pass overflowed it and then jumped backwards at the next scene.
• Identity transfer follows ComfyUI's new RoPE matrix layout.
• Models & Pipeline treats ComfyUI V3 MultiType widgets as widgets rather than required sockets, so a node using them no longer reads as unsatisfied.
• The latent upsampler loads on ComfyUI v0.29.0, which moved it onto DynamicVram and made operations a required argument.

[3.4.0] - 2026-07-24

Added

• Easy Gen (the simplified single-scene UI) reaches near feature-parity with the Cutting Room for the workflows it's meant for:
  • Shortcut/transition JSON library import & export in Settings ▸ Shortcuts (merge or replace) — Easy Gen previously had no way to load a shortcut/transition library at all.
  • The full Engine settings panel (Studio refinement/adjustments/sampler algorithm, Chain Sampler continuity/timing/guidance/decode/experimental, Best-FaceID identity transfer) — restored via the same shared frontend module the Cutting Room uses.
  • A Log button (ComfyUI backend log, same as the Cutting Room's).
  • An explicit Save button, and a Gallery picker to reuse already-uploaded media as the i2v anchor without re-uploading — with a one-click continuity identity pin button right on each thumbnail, mirroring the Cutting Room's Media Bin.
  • An Export… button to download the project as a .funpack_project.json file (e.g. to move a project off a rental GPU box) — the backend route already existed but had no UI.

Changed

• Loop temporal style now works alongside guides. Scenes carrying appended guide frames (carry-i2v-guides, mid-scene guide, JoyAI memory, custom guide stacks) no longer disable the loop roll: only the content region rolls while the pinned guide tail (and the audio memory tail) stays canonical, so guides keep informing every position of the cycle without ever being rendered into it. The i2v anchor was always supported - it rides the denoise mask, which rolls in step with the latent.
• Easy Gen's Studio always runs in Prompt-only mode. Easy Gen has no rating UI, so every Engine setting that's a no-op without a trained refinement key/rated history (refinement key, value guidance, steer mode, reference injection, embed/score/output guidance, taste retrieval, DynaShift, sampler-panel velocity bias/rescue) is hidden there and enforced at the pipeline level, not just the UI — every Easy Gen generate call forces Studio's mode to "Prompt only" and the rating-gated Chain Sampler knobs off via node overrides, so a project configured with these on in the Cutting Room still generates plainly from Easy Gen. A note points to the Cutting Room or the ComfyUI graph for the full learned refiner.
• The "expose to editor" eye buttons in Models & Pipeline are hidden under Easy Gen — there's no per-scene inspector for an exposed control to ever surface in, so they were inert clutter that could look actionable and silently do nothing.

Fixed

• The shortcut autocomplete menu could open off-screen below the prompt field (Easy Gen's prompt box sits in a footer bar at the bottom of the viewport) with no way to see or click a suggestion. Now flips above the field when there isn't room below.
• A node's bypass toggle could silently do nothing: when the pass-through mapping was ambiguous, or the bypassed node's own matching input wasn't wired to anything, the failure was only ever recorded in a report field no frontend code read or displayed — the node stayed fully active with no visible error. Generation now blocks with a clear reason instead of a silent no-op.
• Easy Gen could silently revert Models & Pipeline edits (bypass, exposed widgets, LoRA settings): any later project save sent a stale in-memory copy of the models config back to the server, clobbering whatever was just changed there independently.
• Easy Gen's preview restarted playback (ignoring a pause) on every background health poll or prompt keystroke, since each one rebuilt the video element from scratch.
• A "fetching process...aborted by the user agent" unhandled-rejection error banner could appear when switching between Gallery/generated media in Easy Gen.
• Downloaded generation results always suggested "result.mp4" as the filename regardless of the actual render (no Content-Disposition on the /result route) — the underlying file was already uniquely named per run and already lived in ComfyUI's temp dir, just wasn't surfaced to the browser's save dialog.
• Easy Gen's Generation settings showed a "Seed" field always reading "1" and looking fixed. It was bound to the Project's legacy seed field, which nothing in generation has ever read (the real seed resolution only reads sampler_inputs.seed, already correctly random when unset) — removed the dead field in favor of the working one under Engine ▸ Timing & Seed, which already says so.

[3.3.0] - 2026-07-15

Added

• Loop temporal style - the "loop" option in Temporal style (inert since it was added) now produces seamlessly looping clips, the LTX2.3 equivalent of WanVideoWrapper's Loop Args: on eligible denoise steps the latent is cyclically rolled along time before the model forward and the prediction unrolled after, so the model repeatedly smooths the video's seam (last frame | first frame) as if it were an ordinary interior cut. Video and audio roll by their own frame counts and stay time-aligned; i2v anchor masks roll in step. Near-zero overhead (no extra forward passes). Also honored by the "auto" director when a scene's prompt asks for a loop ("seamless loop", "cycling", "repeating"). On scenes that carry guides (mid-scene guide, carry-i2v-guides, JoyAI memory, custom guide stacks) the roll stays off - guide frames are pinned to absolute positions - and the run log shows temporal_loop_roll(inert: guides pin frames).
• Plateau step-cache (experimental, Engine ▸ Experimental) - reuses the base model's forward across the near-noise plateau steps of the LTX 8-step schedule (sigma >= threshold, default 0.975), skipping 3-4 of 8 forwards per scene for a negative-overhead speedup. Guidance wrappers still post-process every step's prediction.
• Wired widget locks - node widget fields fed by an incoming wire are now locked in the Models & Pipeline editor and show which node feeds them, instead of silently accepting edits that the wire would overwrite on the next run.
• Per-prompt taste direction (taste_nearest_prompt, experimental, Engine ▸ Guidance) — a retrieval layer over the learned taste steering. Instead of pushing Embed guidance / Score slider along the single global liked-direction average, each liked rating now also records a (prompt fingerprint → that run's liked direction) entry, and with this on each scene retrieves the similarity-weighted direction of its nearest rated prompts (a forest prompt pulls what worked on forests, not the mean across every prompt). Non-parametric (a plain ring buffer keyed on pooled conditioning) — no extra model forward, just a cosine lookup + vector mean, and no small value function to collapse into a spurious attractor. Falls back to the global liked direction when nothing rated is close enough. Rides the same refine_v2 clip-state file as liked_dir, so it's removed atomically with the key.

Changed

• Best-FaceID identity transfer (experimental, Engine ▸ Experimental) is now a full native port of ComfyUI-BFSNodes' LTX Identity Transfer overlap+source_phase+ArcFace conditioning, replacing 3.2.1's source-phase-only approximation. The reference face is now injected as separate, non-rendered overlap tokens (never blended into a real frame) instead of tagging Continuity's Identity pin keyframe, plus an optional ArcFace projector channel (identity_projector, id_strength, arcface_mode) for a second identity signal. Still keyed off Continuity's existing Identity pin image — no new picker. Requires the insightface package (added to requirements.txt; downloads the buffalo_l model on first use) only when an ArcFace projector is selected — the overlap tokens alone work without it.
• Frame overlap now carries through anchor changes (hard scene cuts with a new anchor image used to drop the latent overlap; carry_overlap_through_anchor), and an Identity pin configured on an anchor-swap scene resolves again instead of being skipped.

Fixed

• Preview stability (rounds 3-5) - the remaining causes of stale / black / undecodable scene previews, especially "the second half stops loading":
  • /result faststart-remuxes raw ComfyUI renders (VHS writes the MP4 index at the end of the file; browsers cannot seek that) and never serves a video mid-write - a file whose index isn't on disk yet answers 503 until the save finishes, instead of undecodable bytes.
  • A failed remux answers 502 with rate-limited re-attempts instead of permanently falling back to the raw file; temp-dir failures and the loopback fallback path can no longer serve raw index-at-end video either (the fallback also gains Range/seek support).
  • The player materializes video elements only in a window around the playhead instead of one per timeline clip - a montage-scale timeline (100+ clips) used to create 100+ elements and fire ~180 preview-segment encodes at once, starving the browser's per-origin connections (and hitting Chrome's hard 75-media-players-per-page cap), which left the back half of the timeline permanently "loading". Timeline filmstrips load through a small shared queue for the same reason.
  • The monitor shows honest readiness slates ("Video is processing…" / "Video is loading…" / "Preview unavailable") instead of a black or blinking frame, and a long GIL-bound save can no longer exhaust the retry budget.
• Exposed controls silently reverting: the autosave snapshot was taken before edited model values synced back to the project, so an exposed-widget edit could be undone by its own autosave.
• ALG crash on multi-scene runs (tensor truthiness in the alg_latents fallback).

[3.2.1] - 2026-07-06

Added

• Shortcut Ideas bulb (💡 next to prompt fields) is now a habit-aware co-pilot: alongside the existing missing-category browser, it surfaces shortcuts you've historically paired with what's already in the prompt, and what usually follows the previous scene in your own projects — mined on demand from saved project text, no new persistent state.
• Media gallery: one-click continuity pin — a 📌 hover action on image cards sets/unsets the project's Continuity identity pin directly from the Assets gallery, without a trip to Engine settings.
• Shortcut revolver — an opt-in no-repeat cycling mode for multi-replacement shortcuts (sequential or shuffled). A shortcut with several replacements now draws through all of them before repeating, instead of a seeded random pick that could repeat early.
• Guide blur (alg_blur_guides) now stands alone from the sampler's anchor ALG: it works with alg_enabled off, and has its own strength / sigma-threshold controls independent of the anchor's.
• Best-FaceID identity transfer (experimental, Engine ▸ Experimental) — tags the existing Continuity "Identity pin" guide with the source-phase RoPE rotation Best-FaceID-style identity LoRAs were trained on. A native port on the Scene Chain Sampler, no external custom-node dependency; the LoRA itself loads the normal way via Models.

Changed

• Models & Pipeline "+ New node" search collapses to a compact summary once a node is picked, instead of leaving the full results list open above the Values/Wiring panel.
• Prompt autocomplete no longer caps suggestions at 8 — a prefix shared by many shortcuts now shows every match (the menu already scrolls).

Fixed

• Preview stability: the preview-segment ffmpeg re-encode ran synchronously inside the request handler, freezing ComfyUI's event loop and stalling every other streaming response — now backgrounded with per-key locking. Ghost clips (and deep-trimmed lone clips) get proper segments instead of deep-seeking the raw chain file. Segment URLs encode the clip duration so re-trimming a clip can't serve an hour-old cached segment. Pooled preview videos default to preload=metadata (only the current clip and its neighbours preload fully), so a full timeline no longer starves the browser's connection pool mid-playback.

[3.2.0] - 2026-07-05

Added

• Unified macOS-style Settings window (⌘, or the Settings button) — one door to every setting. Icon-rail sidebar that expands on hover, with grouped sections: About FunPack, Editor, Engine, Models & Pipeline, Refinement & Taste, Updates & ComfyUI, Temp Files. The FunPack menu and the old Settings dropdown are gone — no duplicated entries anywhere; the top bar is File / Edit / View / Help / Settings.
• About FunPack — version (now served via /api/git/status from pyproject), commit, branch, and a Software Update deep link.
• Models & Pipeline redesigned around its own sidebar: Linked inputs and Pipeline at the top, "＋ New node", and every configured node with a status dot and hover rename. Adding a node opens a Setup dialog with search where widget values AND output wires / input sources are set before it lands in the pipeline. Link mode is a persistent bar that follows you across node pages. Import ComfyUI Workflow lives here as a header action.
• Engine settings redesigned the same way: always-visible categories (Overview · Studio: Refinement / Adjustments / Sampler algorithm · Chain Sampler: Continuity / Timing & Seed / Guidance / Decode / Experimental) with macOS-style rows and live changed-count badges — no more scrolling a wall of collapsible cards.
• DynaShift (experimental) on the Scene Chain Sampler — negative latent memory: Awful and wrong-appearance ratings bank the run's video latent, and later runs steer x0 away from matched banked frames per late step ("a negative prompt at CFG=1").
• output_guidance on the Scene Chain Sampler — a value function trained on the model's own prediction, nudging sampling toward liked outputs; exposed in Engine ▸ Guidance.
• Value function training: pairwise ranking loss + ensemble disagreement gate.
• Distilled Flow: quality_sharpness unsharp-mask port from Hybrid Euler 2S.

Changed

• Split at playhead actually cuts the clip in place (both halves keep playing their own portion) instead of only adding a marker.
• Settings fields render single-column and shrink correctly, so the window works on small screens; no control can push the layout out of reach anymore.
• Removed block steering and the NVFP4 loader after failed live validation.

Fixed

• Preview player: rapid play/pause could restart the video from the beginning. An aborted stream fetch (or the retry path's cache-busted reload) reset the <video> element to 0 and the playhead followed. The player now stashes the position when the element resets, restores it as soon as metadata is back, and resumes playback if it was playing.
• Sampler: exception-safe per-scene wrapper lifecycle on interrupt + output_guidance gradient calibration.
• Editor: typing no longer interrupted by autosave re-renders; renders no longer become unviewable after transient errors; media thumbnails release their network connections (Chrome's 6-per-origin pool starvation).
• Models "+ Add" requirement buttons doing nothing on an empty project; the disabled-built-in-pipeline state is surfaced instead of hidden.

[3.1.3] - 2026-06-29

Added

• ALG anchor de-staticking (experimental) on the Distilled Flow sampler — adaptive latent guidance that swaps the model's latent image per step (blurred ↔ sharp) to keep i2v anchors from freezing the shot. Opt-in, off by default; alg_blur_guides extends it to guide-attention frames.
• Momentum Guidance (experimental, arXiv:2602.20360) on the Distilled Flow sampler — an EMA-of-velocity smoothing paired with ALG over complementary sigma windows. Opt-in, off by default.
• Bounded Attention (experimental) on the Scene Chain Sampler — reduces attribute bleed between multiple subjects via text cross-attention masking, with exact-boundary per-span encoding and activation/skip logging.
• Auto Montage. Build a trailer-style cut from already-rendered clips; works on generated scenes, not just imported video clips.
• Bypass mode for Models-modal nodes, exposable to the main editor via the existing eye-button / exposed-widget pattern.
• Conditioning Adjustments (per-phrase universal steering) are now exposed in Engine Settings.
• Reconnect to a running generation after a UI reload. If the editor is reloaded mid-run it re-attaches to the in-flight generation in ComfyUI's queue — Generate stays blocked, Interrupt is offered, and the result is recorded onto the right scenes when it finishes.
• Temp files browser (Settings ▸ Temp files) — a media-bin-style view over ComfyUI's temp directory (scene previews and other transient outputs); open any item in a new tab or save it to disk before a restart wipes it.
• Switch Branch / Update FunPack from the welcome screen, so you can swap branches or update without first loading a project.

Changed

• JoyAI-Echo memory strength is no longer capped at 0.5 (now up to 10.0).

Fixed

• Guide scenes crashed on Blackwell (sm_120) GPUs. xformers has no masked (tensor-bias) attention kernel there, so guide/mid-scene-guide runs aborted while plain anchors worked; masked attention is now auto-routed to SDPA.
• Removed a module-level persistent CLIP encode cache that could carry stale conditioning across requests — the UI is the only source of truth.
• Custom shortcut refinement keys are now trained on single-scene runs too, not only multi-scene runs.
• Splitting a clip could leave a phantom/duplicated video transition on the first half.
• A stale scene anchor/render could survive a global-prompt reorder; scene matching is now content-first.
• A previous scene's preview became unwatchable once the next scene generated (temp files now use a unique per-run prefix).
• Audio-track resize could overshoot, and keyboard focus could get stuck on the Start(s) input.

[3.1.2] - 2026-06-28

Fixed

• Cutting Room: splitting a clip could undo itself. An async probe that adopts a render's real encoded file duration could resolve after a split and silently rewrite the first half's length back to the full original — pushing the second half later, which looked like "a full-length scene got appended to the end of the timeline" instead of an in-place split. The probe now carries an edit token that the split bumps, so a stale probe result is discarded instead of overwriting a fresh split.

[3.1.1] - 2026-06-26

Added

• Prompt $variables + global-prompt templates. Define $name variables in the Composer and reference them anywhere in the global prompt or scene text; resolved last (after shortcuts and scene splitting), so existing trigger/split behavior is untouched. Recursive/cyclic references fall back to the literal text instead of erroring. Save the current prompt as a reusable template and load it back from a dropdown above the prompt box.
• JoyAI-Echo memory mode on the Scene Chain Sampler — a cross-shot audio/video memory bank built on guide attention: a video memory bank carries visual identity across shots, paired with an audio memory (protected via a masked tail) and a v2a_grad_scale control over the native video-to-audio attention hook (1.0 = no-op). Toggle in Engine Settings.
• Contrastive-pair FreeSliders. The minus pole of the score-space taste slider is now synthesized from a learned bad_dir, giving a real contrastive pair instead of a single learned direction.
• Scene postfix. Shared text can be appended to every scene, toggleable, with shortcut expansion and key fold-in covered by tests.
• Cycle-guard + smart auto-wire for full-control graphs, so automatic wiring no longer fights manual connections.
• Path-outcome planner (phases 1–2d). Per-scene path-outcome memory with seed-lever-avoidance and Thompson/UCB explore-exploit seed routing, plus conditioning-variant routing — ratings now steer future generations away from disliked/repeated paths.
• BachVid + KV-Lock. Training-free per-key raw K/V identity bank (capture/bless/inject) gated by a variance-of-x0 scheduler that drives injection strength.
• Clip length on the timeline now adopts the real encoded file duration instead of the plan estimate; project-mode scene length honors frames_mode so scenes track project length.

Changed

• Scene boundaries are now passed structurally end-to-end; the generic scene N injection marker is removed entirely, and a malformed/incompatible split can never reach the encoder.
• A corrupt or incompatible saved project now opens with a clear error instead of a 500.
• "Just forget it" (refiner reset) now truly leaves no trace.

Fixed

• Leaked Technique-5 forward hooks could survive a reset/key-delete and cause progressive quality drift across runs; hooks are now tagged and stripped on every run.
• Deleting a refinement key was non-atomic and could orphan its value_fn/blessed banks (still applied after "deletion") and leave behind an invisible keyless Absolute store; deletion is now atomic and the keyless Absolute store is surfaced/clearable in the UI.
• JoyAI memory inputs were inserted in the middle of the sampler's INPUT_TYPES, desyncing widget order in the ComfyUI graph; moved to the end.
• Timeline clip order no longer scrambles after editing the global prompt.

[3.1.0] - 2026-06-21

Added

• Cutting Room rebuilt on an OpenCut-style NLE shell. Three-column layout — Assets (media bin) | Preview (program monitor) | Properties + Timeline — with the prompt-craft tools moved into a Composer floating window (drag, minimize-to-title-bar roll-up, maximize) holding Compose (the global prompt), Shortcuts, Splits, and Files tabs. Layout borrows from OpenCut — credited in the README.
• Prompt autocomplete. Typing in the global prompt or a scene prompt suggests matching shortcut triggers (trigger + replacement + category); Enter/Tab accepts and finishes the trigger with a trailing space, ready for the next. Includes an Add-shortcut picker (browse by category) and works mid-prose, not only after a delimiter. Toggle in Editor settings.
• "Anchor as guide" scene source. The scene image still feeds the pipeline like any anchor (so nodes that need it — e.g. an Image Transform deriving width/height — still get it) and steers the scene from a frame-0 guide, while the i2v node is bypassed so the latent stays empty (text-to-video). Per-scene guide strength; declare your i2v node + the input/state to force in Editor settings → Anchor as guide.
• Mixed mode also attaches the scene's own image as a frame-0 guide, so even the first scene (nothing prior to carry) gets guide attention from its image, on top of the carried prior-scene guides. The Mixed anchor image now also shows an inspector picker (parity with image modes).
• Per-browser Editor settings — prompt autocomplete, shared-anchor toggle, and the Anchor-as-guide i2v bypass config.
• Richer text-overlay styling — bold/italic, alignment, outline, shadow, and a background box.
• Composer file management — Replace/Merge on import, Delete-all, and a Files tab to audit and purge FunPack files on disk.
• Inline clip reorder controls on the timeline; managed Shortcut categories / sub-categories for the Compose tab; the node search/filter is now available in every Add-Model mode (not just "Any node").

Changed

• Canonical scene splitter. split_scenes was rewritten as a single piece-walker shared by every consumer (preview + generation), removing the old offset-projection band-aids and divergent split paths. "Wrong appearance" ratings now act as a consistency anchor.
• Generated renders are immutable and full-length — later plan/trim edits never truncate an existing clip. Timeline cut order is independent of plan order; a clip shows what it was generated with, not live plan text; removing a scene from the plan keeps its generated clip on the timeline. i2v anchors drop on the Plan, not the timeline.
• Exposed controls honor real min/max/step, and autosave no longer eats in-progress input.

Removed

• Removed the FunPack Scene Builder node and the Studio Scene Builder feature (the per-key scene database, saved scenes, the Studio Scene tab, and scene-database wildcard cleanup). It was redundant — helpful on paper but effectively unused, with no way to drive it from the Cutting Room, and it risked interfering with shortcuts. Shortcuts, transition splitting, and refinement keys are unchanged. Studio now always uses the connected positive_prompt.
• Removed the character feature (no UI surface remained) and the "Convert to video / scene" buttons.
• Removed FunPackNormalizingSampler (redundant).

Fixed

• "Reset Studio session" silently wiped keys it never listed. The confirmation read the per-scene scene_refinement_keys (which drops keys via its divergence fallback), while the backend reset wipes the scene-count-independent pool (prompt_scene_shortcut_keys.all_keys — every key fired in the prompt). So a second key you were training could be cleared without ever appearing in the confirmation. The preview now exposes refinement_key_pool (mirrors the backend reset exactly) and the confirmation lists from it, so what's shown is precisely what gets wiped.
• A single-transition "silent" split leaked the trigger word into conditioning.
• Shortcut autocomplete never suggested inside prose prompts; split_scenes dropped a scene when a transition was keyed by expansion; styled text overlays crashed on newer Pillow (float bbox → int).
• The global prompt is now materialized before generate (the screen is the source of truth); the timeline re-renders after prompt edits; stale LoRA/combo slot values coerce to live choices; the Composer "Minimize" rolls the window up in place instead of acting like Close; the autocomplete menu can no longer be orphaned on screen when a panel re-renders.

Notes

FunPack 3.x development focuses on the Cutting Room frontend; pre-3.0 ComfyUI graph nodes remain supported with bugfixes only.

[3.0.2] - 2026-06-15

Fixed

• "Reset Studio session" now lists the keys it will wipe off a fresh preview. The modal read scene_refinement_keys from the cached preview, which can lag the prompt (debounced refresh, or _distributeGlobalPrompt carrying the old keys forward), so a just-added key-bound shortcut was missing from the wipe list (it would only offer to reset default). Arming the reset now awaits a preview refresh before computing the list.
• Exposed project-settings dropdowns now refresh with the model list. A node input exposed to the main editor (e.g. a wired LoRA loader's lora_name + strength) snapshotted its combo options at expose time, so "Refresh model list" updated the live spec inside the Models menu but left the exposed dropdown stale (a newly installed LoRA was detectable in the node settings but not in the project-settings control). Refresh now re-pulls each exposed control's and shared link's combo choices from the freshly loaded spec and persists, and all three refresh entry points (Models modal button, ModelsModal.refresh, menubar "Refresh model list") go through one shared path.
• Reward poisoning from prompt-repair ratings (drift / wrong-character / distorted gens). The Wrong * ratings (appearance/details/action/combos) are prompt-REPAIR signals — "good gen, but the words/identity were off" — yet they were feeding their 0.0/low reward into value-function training AND the keyless Absolute taste store. That stamped visually-good conditioning as a low-value "valley," so the VF ascent (now always the default key, applied to every scene) pushed conditioning away from good regions into drift, and Absolute stored good gens as global bad-taste. Wrong-* now carry skip_value_function: they still drive prompt repair + relative per-key direction/category memory, but no longer train any reward asset. Quality ratings (Missing-*/Perfect/Nailed/Awful) are unchanged.
• Refinement-key attribution rewritten to the simple, correct model — fixes both "every key trained on every clip" AND "custom uploaded keys not detected." The old code re-split the raw prompt independently and compared scene counts, then guessed (all-keys union → cross-training, or, briefly, empty → keys vanished, including single-scene-with-keys). resolve_scene_refinement_keys now just reads each scene: split with the shortcut-aware split_timeline_verbatim (same scenes generation uses) and collect the keys whose shortcuts fired in each scene's text (anchor keys apply to all). No scene-count comparison, no union/empty guessing — a single scene with a custom key is detected, and each scene gets exactly its own keys. Only when the editor's scene count genuinely can't be aligned (advisor/repair restructured the prompt) does it fall back to the project default key.
• Refinement training law: a scene owned by a custom key trains only that key — the project default key is left untouched for that scene and learns solely from scenes with no custom key. Previously every rated scene also trained the default (and, via the union, every other key), which is what bloated keys and cross-contaminated taste.

[3.0.1] - 2026-06-14

Added

Multi-refinement-key support in the Movie Editor. Refinement keys are now a first-class, per-shortcut training signal. (1) The Engine settings → FunPack Studio card has a Refinement key field, so a project can set/wire its own key (previously the key was fixed and unreachable even in full-control mode); it feeds the FunPackRefinementKeyLoader for Studio / Chain Sampler / SaveRefinementLatent. (2) The Shortcuts editor gains a "Use non-default refinement key" checkbox + key-name field — firing that shortcut means its key is being trained. (3) Multi-key per scene: a scene whose prompt fires several shortcuts bound to different keys now has its conditioning steered by each key and averaged/merged into one (one key ⇒ substitute the default; none ⇒ default). Rating that scene trains every participating key. A key counts for a scene only if one of its shortcuts fired in that scene's text (anchor-bound keys count for every scene); attribution falls back to the safe union if the prompt was rewritten or the split diverges.

Per-scene refinement-key preview. The timeline preview now shows, on each scene, the refinement key(s) it will actually steer with before you generate — explicit keys (with (avg) when more than one), or the project default key (greyed, (default)) when no shortcut key fired in that scene. The preview reuses the exact generation-time resolver (resolve_scene_refinement_keys), so what you see matches what runs, including the safe-union fallback.

Session Reset now wipes every key a run trains, not just the default. Because per-scene multi-key learning trains every refinement key whose shortcut fired in the prompt, a Studio Session Reset now clears all of those keys too (project/default key + each non-default key activated by a shortcut), so no stale per-key state survives a reset (FunPackVideoRefinerV2._v2_reset_prompt_keys). The Movie Editor's "Reset Studio session" now confirms first, listing exactly which keys will be wiped: "This action will reset Studio learning for keys: default, key1, key2… To avoid resetting a non-default key, remove the shortcut that activates it from the prompt. Proceed?" — disarming a mis-armed reset does not re-prompt.

Fixed

Shortcut replacement phrases were split on commas, breaking comma-containing replacements. A replacement like "The video rapidly cuts, showing the next view." was torn into two variants at the comma (Movie Editor splitLines and the backend _shortcut_replacements string fallback both split on ,). The editor then showed it across two lines, expansion picked one half at random (so it never matched a comma-form transition trigger), and editing it back to one comma phrase reverted on save because the comma re-split every time. Replacements are now "one per line" only — commas stay inside the phrase; triggers still accept comma-separation as before. Existing torn replacements self-heal on the next edit + save.

Scene prompt text vanished from the timeline after replacing an i2v anchor. Replacing a scene's i2v anchor is a source-only change — it doesn't alter the prompt-preview key or the selection — so commit() replaced state.project with the saved project without re-rendering the timeline, leaving the stale optimistic DOM (the clip's prompt text could read blank/"empty scene" until a reload or regenerate forced a full re-render). Anchor/source changes now flag the commit to re-render from the authoritative saved state once the save lands.

Editor timeline dropped a scene when a transition trigger was also a shortcut. If a scene-cut marker (e.g. qcut, cut) was also a shortcut — common now that cut markers can carry refinement keys (e.g. qcut → "cuts" key) — the lossless splitter expanded the trigger word away before detecting transitions, so the split was lost or misplaced. Classic symptom: "scene 2 removed from the timeline, scene 1 shows scene 2's prompt" until reload. split_timeline_verbatim now also scans the original (verbatim) text for triggers, catching trigger-shortcuts at their true position; genuine shortcut-driven splits (trigger only appears after expansion) and plain triggers are unchanged, and the result still round-trips losslessly.

[3.0.0] - 2026-06-11

Added

FunPack Cutting Room (Movie Editor) — a full browser-based montage editor served from ComfyUI at /funpack/movie. Build multi-scene projects on a real NLE timeline (ruler, proportional clips, trim, split, drag-to-space pauses, crossfades, per-clip effects), preview the whole sequence in a program monitor with seamless scrubbing, and generate or stitch final video without leaving the UI. Includes a Media bin (upload, filter, sort, rename, export), Characters bible with per-scene assign, global prompt editor with lossless split markers, prompt Shortcuts library, Engine settings (FunPack Studio + LTXAV Scene Chain Sampler panels), Models & Pipeline wiring (built-in LTX pipeline or imported ComfyUI workflow), overlay tracks (text/image compositing with WYSIWYG preview), separated and inserted audio lanes, per-scene Refiner ratings, refinement key import/export, ComfyUI log viewer, git-based FunPack update UI, interactive welcome tour with sandbox demo, and first-run pipeline dependency installer (ComfyUI-Manager + LTXVideo / Video Helper Suite / KJNodes packs).

Interactive Guessing — a new Batch Training mode. Instead of varying the seed, it freezes everything (including the noise seed) and sweeps the conditioning's spread (its "sigma") on video channels along a linear ramp across the N rungs — up to amplify toward overbake, down to dampen — so you can see exactly where it breaks. Each rung records its factor. When you rate the ladder, it learns your safe steering ceiling for that key and auto-caps future steering: a soft pre-limit on absolute/relative steer strength plus a hard clamp on the output conditioning's spread (the guarantee). Audio is never touched (reuses the LTXAV video/audio channel split). The Batch Training tab gains Mode (Regular / Interactive Guessing), Direction, Range, and a Learning toggle — Learning off (either mode) means pure generation that teaches nothing. Combined with the new IMAGES output (all batch videos come out the Chain Sampler's IMAGES port), you can now generate N slight variations and just watch them without rating.

FunPack Normalizing Sampler — a new SAMPLER (selectable in Studio alongside Hybrid Euler 2S / Distilled Flow / KSampler) built for distilled few-step LTXAV at CFG=1. Video-only latent normalization counteracts overbaking / oversaturation / colour-drift; audio stays on plain euler. Node FunPackNormalizingSampler + Studio sampler panel.

Changed

The Chain Sampler's IMAGES output now returns every batch video concatenated, not just the last one — for both regular batches and Interactive Guessing.

Latent normalization stacks on the Hybrid and Distilled samplers too. The same video-only anti-overbake normalization from the Normalizing sampler is now an opt-in normalize_strength / normalize_start_sigma on the Hybrid Euler 2S (CONST/RF path) and Distilled Flow samplers. Default off; audio is never touched.

Distilled Flow AB2 ramp — graduated 2nd order. New opt-in toggle that ramps the AB2 contribution linearly from 0→1 across the schedule. No effect at order=1. Exposed on the Distilled Flow sampler node and in Studio's sampler panel.

FunPack development direction (3.0+). New work focuses on the Cutting Room frontend and its pipeline integration. Pre-3.0 ComfyUI graph nodes remain supported but will receive bugfixes only — no major UI reworks on legacy node popups unless a fix requires it.

Fixed

Audio corruption in the FunPack Hybrid Euler 2S and Distilled Flow samplers on LTXAV. AB2 and Heun are now confined to the video stream; audio rides plain 1st-order euler. Distilled Flow's optional s_noise is likewise video-only.

Movie Editor: timeline pause UX (drag clips apart to create pauses, drag flush to remove), preview minibar crash on gap segments, pipeline dependency installer with ComfyUI-Manager bootstrap and cancel, overlay/audio lane stability, multi-scene chain preview offsets, and numerous NLE polish fixes across the dev cycle.

[2.7.8] - 2026-06-04

Added

Made all conditioning steering audio-safe on LTXAV. LTXAV conditions video and audio from two separate text cross-attentions that the model carves out of one conditioning tensor by splitting its channel dim (comfy/ldm/lightricks/av_model.py _prepare_context: torch.split(context, [v_context_dim, a_context_dim], -1) — leading channels → video attn2, trailing → audio audio_attn2). Previously every steer (relative/absolute pull, value-function ascent + search, embed_guidance, the attn2 direction patch, and manual Conditioning Adjust) shifted the whole tensor, corrupting the audio's own text conditioning and degrading audio. Now a shared protect_audio_channels confines every edit to the video channel-slice and restores the audio slice from the unsteered conditioning — effectively "modified conditioning for video, original for audio", with no model patching or extra forward pass. The split is auto-detected from the channel width (7680→3840, 6144→4096) and logs the detected layout once; single-stream LTXV (unrecognised width) is a clean no-op, so video-only models are unaffected. Note: this protects the audio's direct text conditioning (the dominant lever); the per-block cross-modal video_to_audio_attn can still carry a weaker second-order influence from heavily-steered video.

Added an Absolute / Relative steer mode to Refiner V2 and Studio. Relative (the default, unchanged behaviour) is per-prompt: it learns and applies the best conditioning for one specific prompt. Absolute is prompt-agnostic — it accumulates a single global "taste" prior across every rated generation and pulls conditioning toward it regardless of the prompt. In Absolute mode a rating means "this has (or lacks) details I like in general", not prompt adherence: a Perfect means "I love this, give me more of it everywhere", and a low rating means "this is missing the details I like in general". The pull layers two engines — a pooled liked/disliked direction (the learned, automated analogue of the manual Conditioning Adjust phrase shift) and the keyless value function on top — and is applied at the conditioning output (_v2_finalize_conditioning). Both keeps the per-prompt fit and layers the global prior under it; pure Absolute bypasses per-prompt memory entirely. The global store learns from rated runs even with no refinement key wired (it is keyless by design). The Scene Chain sampler's embed_guidance gains an embed_guidance_source (relative / absolute) so the per-step nudge can draw from the global taste direction too. Studio exposes Steer mode + Absolute strength in the refiner panel.

Ported velocity bias + reactive rescue to the FunPack Distilled Flow sampler (sample_funpack_distilled_flow), reusing the same capture/memory/rescue machinery as the Hybrid Euler-2S and LTXAV/RF samplers. The sampler now exposes velocity_bias_mode, velocity_bias_strength, velocity_bias_source, velocity_refinement_key, rescue_mode, rescue_threshold, and rescue_strength, and these are wired into Studio's Distilled Flow sampler config panel (with the same blank/default velocity-key → wired-refinement-key fallback the Hybrid panel uses). Steering is applied magnitude-preservingly and, on packed LTXAV latents, confined to the video stream (audio-safe). Note: few-step distilled schedules may only land on one or two velocity targets, so apply/rescue fire less often than on an 8-step run; they no-op cleanly when no target matches. With everything off the sampler is byte-identical to the previous deterministic ODE.

[2.7.7] - 2026-06-03

Added

Added Batch Training — a controlled-batch RLHF workflow built around the principle that the cleanest learning signal comes from rating several generations that differ in exactly one thing. When a batch is active, the Scene Chain sampler runs the chain N times with everything frozen except the noise seed, producing N directly-comparable videos. Studio then shows a rating panel where every variant is scored, and on submit the value function trains on the batch's shared (frozen) conditioning with each variant's reward — same conditioning, N rewards, which is the variance-reduced comparison signal. Spans the full pipeline: the engine (controlled N-run mode on the sampler), a Studio variant producer that packs the variants into conditioning, the /funpack/batch server routes and rating window, and the value-function intake on submit. Phase 3b adds deeper learning from batch ratings via axis and direction memory. Batches live in ComfyUI's temp directory and are wiped on restart.

Added a reactive in-flight rescue system to the Hybrid Euler-2S sampler. During sampling the trajectory is compared against conditioning-clustered memory of past good and bad runs; when a step drifts toward a known-bad trajectory it is nudged back. Rescue is rating-gated (separate good/bad trajectory banks, learned only from rated runs), prompt-aware (clustered by conditioning so the right memory is consulted), runs in both sampler phases, and persists its trajectory banks to disk. The full feature set was also ported to the LTXAV rectified-flow (CONST-model) path via a sampler mirror, so velocity capture and rescue now work for LTXV/LTXAV, not just the discrete-step models.

Added Monte Carlo conditioning search after the value-function gradient ascent, plus a VF final gate, so the conditioning chosen for generation is the best of a sampled neighbourhood rather than the raw ascent endpoint.

Added a Studio Variability macro and an active-feature readout for the Hybrid sampler, summarising which steering features (velocity bias, rescue, embed guidance, value guidance) are live for the current configuration.

Changed

Unlocked velocity_bias_strength from a 0.35 cap to 3.0 for deliberate creative action injection. Velocity bias is an artistic / action-injection control (and carries an emergent scene-cut prior — cuts survive the trajectory mean and get reintroduced unprompted); it is not a consistency tool.

Refiner V2 now always trains the value function; value_guidance only gates whether the learned reward is applied during sampling, and defaults on. This means rated runs keep teaching the value function even when guidance is off.

Decoupled eta_final decay from the quality boundary — it now anchors to schedule progress instead, so ancestral-noise decay tracks the sampling timeline rather than a quality threshold.

Fixed

Audio-safe LTXAV sampling: ancestral noise and trajectory steering are now applied to the video latent only, never the audio latent, preventing the joint-attention audio corruption that steering on the combined tensor caused.

Velocity-bias anti-softening: the bias is now applied magnitude-preservingly with sigma-decay and a quality-sharpness term, and sourced from the nearest trajectory cluster, so it injects motion without washing out detail.

Audit fixes: corrected a velocity commit key mismatch, sparse rescue targets, and a redundant aspect-bucket computation. Batch Training fixes: distinct per-variant seeds in both the split-scene and single-scene paths (identical seeds were producing identical videos), correct activation alongside split_by_transitions, the node rating is ignored while a batch is in progress, and the in-node panel was de-duplicated from the Studio Refiner tab.

[2.7.6] - 2026-05-30

Added

Added pre-generation conditioning ascent driven by the value function. Before sampling, the positive conditioning is moved by gradient ascent toward higher predicted reward, in both the single and scene-split Refiner V2 output paths. Displacement is capped to prevent reward hacking (the value function will otherwise push conditioning into degenerate high-score regions). The value function can be exported and imported from Studio and is cleared on session reset.

Added VF-driven conditioning shaping: confidence scaling and gradient-aligned phrase boosting in conditioning memory, attention-weight accumulation, and VF-driven temperature, so the learned reward influences which phrases and attention weights are emphasised.

Added concept-in-context conditioning guidance and concept-pair bad-direction repulsion — conditioning is nudged toward liked concepts in context and away from concept pairs learned to be bad.

Added a motion floor to embed guidance: when temporal variance falls below a threshold the guidance auto-boosts it, fighting static/frozen output. Activation is logged with step, sigma, and variance ratio.

Reworked the rating UI into a rating picker popup. Added a Wrong action and an explicit quality rating, a Nailed it rating (prompt-adherence positive, weaker than Perfect), and replaced the standalone Loved it rating with a per-option heart modifier (an axis-blind quality endorsement layered on top of any rating). The heart is disabled for quality-degraded ratings and for Awful.

Changed

Disabled perfect_freeze — prompt changes made after a Perfect rating are now respected instead of the conditioning being frozen.

Removed

Removed the experimental latent value function (VGG-Flow-style per-step latent steering). It was added in this cycle but pulled: the LTX token format does not expose the latent in a form the per-step steering could act on.

Fixed

Fixed an UnboundLocalError (eff_key not yet assigned when the value function loads), value-function loading needing inference_mode(False), several rating-picker bugs (stale outside-click listener re-opening the picker, active state not updating on option change, clicks landing on a canvas button widget rather than a DOM element), and Save Refinement Latent not executing.

[2.7.5] - 2026-05-29

Added

Added embed guidance — a per-step nudge of the conditioning toward the learned "liked-quality" direction during sampling. Near-free overhead; requires a wired refinement_key_input with enough liked generations to have formed a direction. Exposed as embed_guidance / embed_guidance_strength on the Scene Chain sampler.

Added an online value function for reward-guided sampling — a small MLP trained on rated generations that predicts reward and can steer sampling toward it. Required several fixes to train and run gradients inside ComfyUI's inference_mode execution context.

Added mid-scene guide (mid_scene_guide / mid_scene_guide_strength), replacing the broken self_consistency feature. It uses the LTX guide-attention mechanism rather than post-block hidden-state injection (which corrupted audio through joint attention). At ~0.25–0.3 strength it preserves static-element layout across a scene; capped at 0.5.

Added a vision-conditioning toggle to the Studio popup, and moved reference-image conditioning into Studio, where the source image already lives.

Changed

Removed the predefined transition-phrase list — scene splitting is now fully driven by user-defined and auto-detected transitions.

Removed

Removed the per-scene vision re-encoding and clip input from the Scene Chain sampler, and reverted the per-scene reference-image encoding in Studio (both added in 2.7.4). They were unstable in practice.

Removed self_consistency (corrupted audio via joint attention — superseded by mid_scene_guide), the i2i_scene_cut feature, and dead guide-keyframe / guide-conditioning code paths.

Fixed

Fixed i2i_strength inversion (higher now means stronger reference influence) and replaced the 1-frame i2i anchor with 2-frame i2v anchor generation for hard cuts.

[2.7.4] - 2026-05-28

Added

Added K/V in-context conditioning for LTXAV identity blocks during i2v generation. Reference hidden states are captured at the start of each scene's denoising pass and prepended as extra attention tokens to the identity-formation blocks ([14, 20, 21, 30, 33]). This forces the model to attend to the reference character's appearance during every self-attention step in those blocks. Result: strong character consistency across scene cuts, view changes, and orientation changes with no LoRAs required.

Added Gemma3 vision prompting. When source_image is connected to Studio and the CLIP was loaded with a Gemma3-12B checkpoint via DualCLIPLoader, Studio automatically encodes the reference image through the built-in SigLIP vision encoder and feeds the resulting vision tokens into the text conditioning. This means the model conditions on both the prompt text and the actual pixel content of the reference frame. No extra node is required — DualCLIPLoader already loads the vision weights when present in the checkpoint.

Added per-scene vision re-encoding to FunPack LTXAV Scene Chain Sampler. Connect an optional clip input. After each scene is sampled, the next scene's conditioning is re-encoded using the previous scene's decoded last frame as vision context. This gives identical scene texts genuinely different conditioning based on runtime-generated content, so the model knows what state it came from when building the next scene.

Added duplicate scene text differentiation. When two or more scenes share identical text, the second and subsequent occurrences are encoded with a "Returning to an earlier scene: " prefix. The original text is preserved in metadata for logging. This breaks the shared conditioning cache entry so the model receives distinct input for each occurrence.

Improved

Reworked frame_overlap=0 soft continuation: the previous scene's last 4 frames are now prepended with mask 0.4 (partial denoising) instead of 1 frame at mask 0.0 (fully pinned). The model receives temporal context from the previous scene while retaining enough denoising freedom to commit to orientation and pose changes directed by the text prompt.

Fixed

Fixed anchor text punctuation when joining scene segments. If the character description ends with ., !, ?, or ,, a plain space is used as the separator instead of injecting a redundant , . Previously a description ending with a full stop produced "description., scene text".

Removed

Removed FunPackGemmaVision node. It was redundant: DualCLIPLoader with a Gemma3-12B checkpoint already loads vision_model and multi_modal_projector weights through the normal load_state_dict path, so the manual weight injection the node performed was a second load of the same file. The vision capability check now detects the attributes directly instead of relying on an injected flag.

[2.7.3] - 2026-05-27

Fixed

Fixed carry_i2v_guides soft continuation when frame_overlap=0: the anchor mask is now fully pinned (0.0) to prevent denoising from disturbing guide tokens.

Notes

Using frame_overlap=0 together with carry_i2v_guides=True is confirmed to produce bad results and is not recommended. Both parameters now warn about this in their tooltips and in the sampler documentation. Use this combination only for deliberate testing.

[2.7.2] - 2026-05-22

Added

Added Shortcuts system. Activation phrases in the prompt are replaced with full cinematic descriptions before encoding. Multiple replacement options are randomly picked per seed. Empty replacement removes the matched phrase entirely. Longer phrases always win over shorter overlapping triggers. Managed in the new Studio Shortcuts tab with Add/Save/Delete/Import/Export.

Added Transitions system. User-defined transition phrases extend the built-in split list. Custom entries support a placement override - start (transition opens the new segment), end (transition closes the previous segment), or silent (split happens but the phrase is stripped from output entirely). Managed in the new Studio Transitions tab.

Added global Transition placement setting in Studio Refiner tab (start / end / silent), with per-entry override on each custom transition entry.

Fixed

Removed all single-word temporal markers (next, suddenly, later, finally, etc.) from the built-in transition phrase list. They caused false splits on normal prose.

Fixed _GENERIC_SCENE_LABEL_PATTERN matching "scene proceeds", "scene features", "scene shows" and similar noun-verb constructions as scene labels.

Fixed dangling trailing transition segments (prompt ending with "...cuts to the") being kept as near-empty scenes.

Fixed stray comma artifacts after article words when transition phrase merging occurs.

Fixed custom transition triggers ending with punctuation (e.g. "Scene cut.") not being detected due to a misplaced word boundary.

[2.7.1] - 2026-05-21

Added

Added successful seed memory for FunPack Studio. When the seed output is connected, Perfect and Loved it ratings store the previous run's sampler seed under the active refinement key. Future runs occasionally reuse concept-matched successful seeds while keeping normal fresh seeds as the default path.

Added per-scene seed metadata for prompt-split mode. Studio and Refiner V2 keep the public seed socket as a single integer, while each detected scene conditioning entry can carry its own funpack_scene_seed for the Scene Chain sampler.

Added use_same_seed to FunPack LTXAV Scene Chain Sampler. When enabled, every scene uses the first provided scene seed or the base seed. When disabled, each scene uses scene seed metadata or falls back to seed + scene_index.

Changed

Made Scene Chain i2v guide carry opt-in. The default path no longer appends protected i2v frames as hidden guide tokens, preserving cleaner scene cuts unless the experimental option is deliberately enabled.

Fixed

Fixed compact i2v guide masks failing to concatenate with full spatial Scene Chain masks by broadcasting guide masks to the chunk mask shape.

Updated README release notes and added the Intent section.

[2.7.0] - 2026-05-21

Added

Added FunPack LTXAV Scene Chain Sampler for split-scene LTXV/LTXAV continuation in one ComfyUI run. It consumes multi-entry positive conditioning from FunPack Studio or FunPack Video Refiner V2, samples one scene chunk per conditioning entry, increments the seed per scene, preserves overlap from the previous chunk, and blends/appends chunks in latent space.

Added support for plain LTXV video latents and nested LTXAV video/audio latents in the Scene Chain sampler. For nested AV latents, video and audio tensors are continued together, with audio overlap derived from the video/audio latent length ratio.

Added broad order-only scene splitting for split_by_transitions. Scene labels such as scene ten, scene -999999, and scene minus infinity are transition cues, but their written labels never affect scene numbering or order.

Expanded transition phrase detection with scene progression, camera shift, zoom, final shot, and final transition phrases.

Changed

split_by_transitions=True now returns one conditioning entry per detected scene through the existing modified_positive output. No new Refiner V2 or Studio output sockets were added.

The text before the first transition is treated as a shared character/global anchor and prepended to every detected scene conditioning. This is intended to improve character consistency across generated chunks.

Removed the hard 8-scene cap from Refiner V2 split output and Scene Chain sampler execution. max_scenes still defaults to 8, but users can raise it for longer chains.

Standalone then is no longer a transition trigger. More specific phrases such as and then, camera transitions, scene labels, and explicit cut/transition language still split scenes.

Fixed

Fixed Studio's Scene Builder mode dropdown not refreshing the active tab immediately after selecting a new mode.

Warning

FunPack LTXAV Scene Chain Sampler is resource heavy. Long chains create large final latents and may run out of memory during VAE Decode even when sampling succeeds. Start with short scenes and a modest max_scenes, then increase carefully.

[2.6.0] - 2026-05-16

Added

Added FunPack Studio - a single node that replaces the typical chain of Refinement Key Loader, Scene Builder, Apply LoRA Weights, LoRA Loader, Video Refiner V2, and Conditioning Adjust with a tabbed popup editor. All settings are managed inside the popup; only the rating widget and Open Studio button are visible on the node face.

Studio inputs (in order): model, clip, advisor_clip, positive_conditioning, negative_conditioning, clip_vision_output, source_image, lora_stack, positive_prompt, negative_prompt, user_intent_prompt, feedback_prompt, refinement_key_input.

Studio outputs (in order): model (LoRAs applied + attn2 direction patch), modified_positive, negative (encoded from negative_prompt or passed through from negative_conditioning), seed (for wiring to sampler), high_pass_sampler, high_pass_sigmas, low_pass_sampler, low_pass_sigmas, loss_graph, status, training_info, encoded_prompts.

Studio popup tabs:

• Session: refinement key management and Scene Builder mode selector (Pass-through / Manual / Auto / Learning).
• Scene: scene preset load and save, phrase bank from session memory, positive prompt composer.
• Refiner: all Refiner V2 settings including negative prompt field, feedback, and intent override. Shows a banner and disables the intent field when Scene Builder is active.
• Advisor: enable/configure an internal HuggingFace CausalLM advisor. Uses the same model cache as the standalone Advisor LLM node.
• LoRA: full LoRA pipeline - session weight suggestions are read first, then LoRAs are applied to model and CLIP, then the direction patch is applied on top. Supports model type (ltx2/wan) and per-block settings.
• Sampler: configure Hybrid Euler 2S, Distilled Flow, or any KSampler for the high-pass and low-pass outputs independently. Sigma schedules entered as comma-separated floats.
• Adjustments: phrase-level conditioning adjustments with session phrase bank.

Three text inputs (refinement_key, feedback_prompt, user_intent_prompt) have override toggles: when off, connected inputs win; when on, popup values win.

The popup remembers its last active tab per node via localStorage. All field changes auto-save to the node widget after 600ms, so settings survive page refresh without requiring Close to be clicked.

Added negative_prompt encoding to Studio: when no pre-encoded negative_conditioning is connected, Studio encodes the negative_prompt text via CLIP internally, removing the need for an external CLIPTextEncode node.

Added /funpack/available_loras and /funpack/phrase_memory backend endpoints used by Studio's LoRA picker and phrase banks.

Added FunPackConditioningAdjust standalone node for phrase-level conditioning adjustments. Encodes each phrase via CLIP, computes a unit-norm direction from the base conditioning, and applies it at user-set strength. Positive pushes toward the phrase, negative pushes away. Popup editor with session phrase bank.

Added seed output to FunPack Video Refiner V2 (via optional _seed parameter) so Studio can generate a seed and expose it as an output for wiring to samplers.

Fixed

Fixed FunPackAdvisorLLM tokenize failing with AttributeError when apply_chat_template returns a BatchEncoding instead of a plain tensor in newer transformers versions. Explicit hasattr(result, "input_ids") check now handles both return types.

Fixed advisor generation producing no output for Qwen3 and other chain-of-thought models: thinking tokens (<think>...</think>) were not being stripped because they are special token IDs that disappear with skip_special_tokens=True. Switched to decoding with skip_special_tokens=False then stripping thinking blocks with regex. Truncated thinking blocks (no closing tag, token budget exhausted) are also stripped.

Fixed FunPackAdvisorLLM attention mask warning and erratic output (echoed prompts, missing spaces) caused by pad_token == eos_token without an explicit mask. Now passes torch.ones_like(input_ids) as attention mask to all generate calls.

Fixed FunPackConditioningAdjust adjustments not applying for LTX/Gemma3 conditioning: the node was reading pooled_output which is None for T5-based encoders. Now uses conditioning.mean(dim=(0,1)) on the sequence tensor, matching how V2 handles conditioning internally.

Fixed Refiner V2 advisor diagnostic not being generated in Full mode: the LLM analysis prompt was asking for session-wide pattern recognition when no history existed yet. Now adapts - uses a simple per-run analysis on early sessions and session pattern analysis when history exists. Added a rule-based fallback so diagnostic history always accumulates even when the LLM produces empty output.

Fixed perfect_repair_phrases and _v2_emphasized_prompt injecting phrases regardless of the prompt_repair toggle. Both are now gated behind prompt_repair=False.

Removed the Perfect-rating advisor gate. The advisor previously skipped both analysis and repair when rating was Perfect and no text feedback was provided. Perfect is not a ceiling - the advisor now runs normally for Perfect ratings.

[2.5.3] - 2026-05-16

Fixed

Fixed FunPackAdvisorLLM wrapper not triggering advisor generation. Qwen3 and other chain-of-thought models emit <think>...</think> blocks that were not stripped, causing the parsed repaired prompt to contain reasoning text. This made body-similarity validation reject the result as "too far from intent." Fixed by stripping thinking blocks in decode. Also added enable_thinking kwarg support in tokenize (with TypeError fallback for models that don't support it) and expanded max_new_tokens by 2048 when thinking mode is active so the reasoning budget does not crowd out the actual response. Fixed _v2_text_semantic_similarity returning 0.0 for generation-only clips that have no encode_from_tokens_scheduled - these now return 1.0 (skip the semantic gate) instead of causing spurious rejections. Fixed FunPackAdvisorLLM missing from the standalone import block in __init__.py, which broke the test suite.

Removed Perfect-rating advisor gate. The advisor was silently skipping both the analysis pass and the repair pass whenever the rating was Perfect and no text feedback was provided, even when the user had an active advisor mode. Perfect is not a ceiling - if the user provides feedback_prompt, it must be honored regardless of rating. The only remaining guard is the allow_prompt_change check (Learning mode).

[2.5.1] - 2026-05-16

Added

Added FunPackAdvisorLLM node. Loads any HuggingFace CausalLM (including sharded checkpoints) as an advisor for Refiner V2. Connect the output to advisor_clip. Model is cached after first load so subsequent runs do not reload. Also fully compatible with the built-in TextGenerate and TextGenerateLTX2Prompt nodes - supports skip_template, min_p, presence_penalty, and progressive fallback for unsupported generation parameters.

Added _v2_direction_readout to training_info Adaptation section. Shows each direction memory slot in plain language: run count, magnitude, whether it is in direction mode or lerp fallback, and the role each axis is playing this run.

Changed

Advisor prompt format rewritten to natural language. Both the repair and analysis user messages now read as plain enhancement requests rather than structured field-value pairs. Works with enhancement-type models (Sulphur, Qwen prompt enhancers) as well as instruction-following models. System prompt reduced to one sentence.

Direction-based conditioning now uses max_new_tokens instead of max_length in the FunPackAdvisorLLM wrapper so prompt length does not eat into the generation budget.

Model patch status expanded to show which direction slots are active with run counts and which phrase texts are being emphasized in cross-attention.

Adaptation status block rewritten to multi-line readable format showing strength, reward trend, streak, per-slot mode (direction vs lerp fallback), and axis adjustments applied this run.

Fixed

Fixed _v2_generate_advisor_text returning None when layer 1 tokenization succeeded - generate/decode/return were inside the except TypeError block so they only ran when layer 1 failed.

Fixed session reset not clearing intent_expansion_memory, session_source_mean_count, liked_dir, and bad_dir - these fields were missing from _v2_empty_state and survived reset via setdefault.

Fixed advisor repetition loops: repetition_penalty raised from 1.05 to 1.3, added no_repeat_ngram_size=5, temperature raised from 0.5 to 0.7.

Fixed system prompt bleeding into advisor output by splitting prompts into (system, user) tuples and applying the model's native chat template. Three-layer fallback: native system_prompt kwarg, manual apply_chat_template via BFS, flat string with completion anchor.

Added persistent cross-run encode cache (_V2_PERSISTENT_ENCODE_CACHE, 4096 entry cap) so phrase encodings are not recomputed every run when CLIP and text are unchanged.

[2.5.0] - 2026-05-15

Added

Added a CLIP text-generation advisor to FunPack Video Refiner V2. The advisor runs two sequential passes: an analysis pass that identifies what specifically needs to change in the suggested prompt, followed by a repair pass that applies those findings. Both passes see the current suggested prompt, user intent, previous prompt, memory suggestions, and the full feedback history.

Added advisor_clip input to use a separate generative CLIP/Gemma model for the advisor while the main clip continues handling encoding and similarity checks.

Added advisor_mode dropdown: Off, Only diagnostics, Only prompt, Full. In Full mode both analysis and repair passes run. In Only diagnostics mode only the analysis pass runs and its finding is stored in feedback history for the next run. In Only prompt mode only the repair pass runs silently.

Added feedback_prompt optional input. When connected, the user's natural-language description of what was wrong is placed first in both advisor passes, with the system instructed to follow it exactly and override all other repair logic.

Added persistent advisor_feedback_history stored in V2 session state. Up to ten past feedback entries accumulate across runs, each labelled with the corresponding rating: Missing action: he was supposed to hold her hand not her head. Advisor-generated diagnostics from Only diagnostics runs are stored as Advisor note: entries so they carry forward into subsequent Full runs.

Added Prompt only execution mode to FunPack Video Refiner V2. All prompt shaping runs as normal but conditioning vectors are passed through unchanged. Learning still applies. Useful when conditioning adaptation should be paused while prompt refinement continues.

Added prompt_repair boolean input to FunPack Video Refiner V2 (default on). Turning it off disables the rule-based phrase injection from phrase memory and passes no repair candidates to the advisor. Useful early in a session before enough context has been built.

Added encoded_prompts STRING output to FunPack Video Refiner V2. When the advisor ran and produced a suggestion, the output includes up to four labelled sections: Positive prompt (what was encoded), Advisor suggestion (applied/rejected) (what the advisor generated), Advisor analysis (the diagnostic text from the analysis pass), and Pre-advisor prompt (the prompt before the advisor rewrote it).

Added eta_final parameter to FunPack Hybrid Euler 2S Sampler. When set below eta, ancestral noise strength decays linearly toward this value as sigma approaches the quality phase boundary, smoothing the transition into deterministic refinement. Default 1.0 preserves existing behaviour.

Changed

Replaced the Refiner V2 advisor system prompt with a structured repair format. The advisor now receives four explicit variables — ORIGINAL_USER_INTENT, LAST_PROMPT, RATING, and OPTIONAL_NOTE — and is instructed to rewrite LAST_PROMPT to fix the specific failure described by the rating. Memory suggestions, feedback history, and analysis context are folded into OPTIONAL_NOTE. The repair pass for Only prompt mode outputs a plain prompt string with no labels, matching the instruction to output only the final text.

Removed negative_prompt input and modified_negative conditioning output from FunPack Video Refiner V2. Negative conditioning has no effect at CFG=1.0 with NAG guidance and added a redundant AI generation call in every mode. Both the rule-based negative repair and the negative advisor pass are removed.

Increased advisor token budget: repair pass 800 → 1600 tokens, analysis pass fixed at 1200 tokens (was repair // 2 = 400). The analysis limit is now independent of the repair limit so it does not shrink if the repair budget changes.

FunPack Hybrid Euler 2S Sampler early phase now uses an order-2 ancestral denoised extrapolation (Adams-Bashforth 2-step) in addition to the existing Euler-A update. The previous step's denoised estimate is used to extrapolate a better score direction at zero extra model-call cost. The state resets after any motion pulse.

FunPack Hybrid Euler 2S Sampler quality phase now uses a progressive correction_blend: the first half of quality steps use a single-eval Euler ODE pass; the second half use the configured 2S correction. This reduces model calls in the quality phase while concentrating the expensive correction where sigma is lowest and it has the most impact.

Fixed

Fixed Only prompt advisor mode running two AI generation calls per invocation (one for positive repair, one for the now-removed negative advisor), causing each run to take twice as long.

Fixed encoded_prompts always showing only Positive prompt: regardless of advisor activity. The final return path was calling _v2_encoded_prompts_output without the advisor keyword arguments.

Fixed Refiner V2 advisor generation: do_sample was False, forcing greedy decoding and silently ignoring temperature, top_k, top_p, and all sampling parameters. The model was always producing its highest-probability default output regardless of instructions or feedback. Changed to do_sample=True with temperature 0.5.

Fixed advisor validation silently rejecting valid feedback-driven repairs: the intent-distance check and protected-category checks now bypass when feedback_prompt is connected, allowing the advisor to implement what the user explicitly requested.

Fixed _v2_find_perfect_example_for_intent accessing a field (loved_delta_sources) that was never written. It now correctly reads from perfect_anchors and loved_variants.

Fixed _v2_update_streaks updating conditioning strength signals (avg_reward_ema, good_streak, bad_streak) in Prompt only mode, which contaminated conditioning adaptation for subsequent Refine runs. Rating and axis labels still update for repair continuity.

Fixed advisor rating label on first run or session reset: was forwarding the user's rating widget value even when there was no previous output to apply it to. Now passes "No previous output (first run or session reset)" when has_previous_run is false.

[2.4.2] - 2026-05-15

Added

Added Learning mode to FunPack Video Refiner V2. It still observes prompts, conditioning, ratings, phrase memory, and diagnostics, but passes positive and negative prompt conditioning through without prompt repair, Lucky composition, wildcard cleanup, or conditioning-vector adaptation.

Fixed

Fixed FunPack Scene Builder mode handling so the live Mode widget stays independent from the selected saved scene, including queue-time Learning and Auto behavior.

Fixed FunPack Scene Builder rich prompt editing so the caret can move past a final inline phrase chip with the mouse or right arrow key.

[2.4.1] - 2026-05-14

Fixed

Improved FunPack Scene Builder database rows so long words and phrases show their full text as a hover hint, and double-click editing opens a wider multiline field with explicit OK/Cancel buttons.

[2.4.0] - 2026-05-14

Added

Added FunPack Scene Builder, a scene preset node that replaces FunPack Template Manager. It collects universal prompt phrase memory, lets users manually assign positive/negative scene phrases, passes the current LoRA stack through unchanged, and can auto-apply a saved scene from an intent prompt match.

Simplified FunPack Scene Builder so prompt and intent text are connection-only inputs, removed model-mode and per-block controls, and outputs only scene prompt data plus the pass-through FunPack LoRA stack instead of conditioning.

Added Learning mode to FunPack Scene Builder; it records connected prompt phrases into the selected refinement key's scene memory while passing positive prompt, negative prompt, and LoRA stack through unchanged. Refiner reset clears conditioning-delta learning while preserving the refinement key's Scene Builder memory.

Redesigned FunPack Scene Builder as a compact button-driven node with centered editor menus for scene name, mode, aliases, Positive Prompt, Negative Prompt, and Database controls. First use asks for a scene name before editing, connected prompts now teach useful words as well as phrase chunks, the editor refreshes the selected refinement-key database before opening, prompt editors highlight already-used chips, database words can be double-clicked for inline editing, and wildcard random choice is now a checkbox for adjacent entries instead of a text group.

Added searchable LoRA picking to FunPack Apply LoRA Weights. The compact row UI remains the primary workflow, and saved workflows still serialize through the existing lora_list JSON value.

Added optional clip_vision_output, source_image, and negative_prompt inputs to FunPack Video Refiner V2.

Added a final modified_negative conditioning output to FunPack Video Refiner V2. When negative repair has prompt text to encode and CLIP is connected, the node returns repaired negative conditioning; otherwise it returns an empty conditioning list.

Added advisory V2 vision context storage for source image dimensions, aspect ratio bucket, image fingerprint, CLIP Vision tensor summaries, and changed-image detection. Vision context is diagnostic only and is not blended into positive conditioning.

Added experimental early velocity bias capture/application controls to FunPack Hybrid Euler 2S Sampler, defaulting off.

Changed

Removed public registration for FunPack Template Manager. Use FunPack Scene Builder for new scene/preset workflows.

Updated V2 prompt repair so repaired phrases preserve stopwords and phrase text while still using filtered semantic tokens for matching and categorization.

Reduced repeated Refiner V2 CLIP model calls by caching category and phrase encodes within each run.

Updated negative repair to persist poorly rated or wrong-context tags and append them to future negative prompts before encoding negative conditioning.

[2.3.3] - 2026-05-08

Fixed

Fixed Refiner V2 so CLIP and pre-encoded positive_conditioning can both be optional inputs. When CLIP is connected, V2 keeps owning prompt encoding as before. When CLIP is not connected but positive_conditioning is connected, V2 accepts the finished Gemma3/LTX2 conditioning, uses the prompt for analysis, and loads only the Gemma3 tokenizer.

[2.3.2] - 2026-05-08

Added

Added Refiner V2 original-intent alignment memory. When user_intent_prompt stays the same but an enhancer produces different positive_prompt variants, the refiner now remembers intent-enhance pairs, which variants rated well, which original-intent phrases were omitted, and which enhancer-only phrases were rejected.

Fixed

Fixed Refiner V2 so learned original-intent omissions can be restored on later runs, while repeatedly rejected enhancer-only additions can be removed before encoding. Rejected enhancer-only full words and adjacent word pairs are stored as omit evidence for that original intent.

[2.3.1] - 2026-05-08

Fixed

Fixed Refiner V2 Prompt Repair so missing/wrong ratings only repair from the current prompt or explicit user intent, instead of pulling unrelated learned favorite actions, details, quality cues, camera moves, or styles from memory.

Fixed Prompt Repair memory matching so the same word with different neighboring prompt context is treated as different evidence.

Fixed vague raw user intent handling so prompts like Figure it out let the enhanced positive_prompt drive repair matching when available.

[2.3.0] - 2026-05-08

Added

Added Wrong appearance rating to FunPack Video Refiner V2 for outputs contaminated by remembered clothing, character, subject, or background concepts.

Added FunPack Refinement Key Loader, with a selectable key dropdown, create-on-load behavior, and browser-side JSON import/export buttons.

Added a Discord-friendly Refiner V2 quick guide for new users.

Changed

Updated Refiner V2 Prompt Repair so it only auto-adds safe repair concepts such as action, camera, details, quality, and style. Appearance, subject/character, and environment/background concepts are now blocked from Prompt Repair.

Updated I'm Feeling Lucky in Refiner V2 so appearance, subject/character, and environment/background memory is not auto-injected unless the user explicitly includes that phrase in the current prompt.

Updated legacy Void/Lucky token-bank selection to skip appearance, subject/character, and environment/background tokens.

Updated Refiner V2 and FunPack Apply LoRA Weights so both can accept a linked refinement key from FunPack Refinement Key Loader.

Fixed

Fixed appearance bleed-over where highly liked clothing or character tags could reappear in unrelated image-to-video prompts.

[2.2.1] - 2026-05-07

Fixed

Fixed FunPack Video Refiner V2 prompt phrase categorization so environment and appearance descriptions are not pulled into action learning by generic -ing or -ed words.

Updated Refiner V2 category similarity blending so CLIP category comparisons only help uncertain phrases instead of overriding strong local action, camera, appearance, environment, quality, or detail anchors.

Fixed FunPack Video Refiner V2 so prompt-enhancer refusal text like "I'm sorry, I cannot help..." is passed through without being saved into prompt history, phrase memory, or future learning targets.

Improved FunPack Video Refiner V2 training data output with clearer sections and extra line breaks for run state, learning, prompt analysis, adaptation, guidance, and LoRA diagnostics.

Updated FunPack Video Refiner V2 to remember liked action/detail phrase clusters with their neighbors and use those ordered clusters before weaker ngram or token memory when repairing missing axes.

Added Wrong details, Wrong action, and Wrong details + action ratings for good-looking videos that do not match the requested intent; these preserve satisfied quality/composition signals while marking the mismatched action/detail context for repair.

[2.2.0] - 2026-05-07

Added

Added FunPack Video Refiner V2, a simplified prompt-owned refiner that accepts positive_prompt and a connected CLIP, owns prompt encoding internally, learns from ratings, and returns refined positive conditioning plus diagnostics.

Added FunPack Template Manager, a preset node for storing prompts, activation words, refinement keys, sigma schedules, and FunPack LoRA stacks with import/export support.

Added I'm Feeling Lucky mode to FunPack Video Refiner V2. It works as a preference composer that can inject learned user-preferred actions, camera moves, details, and styles even when the current prompt is vague.

Changed

Updated LTX per-block LoRA loading so supported stacks now compare LoRA block fingerprints across the whole stack and apply type-aware conflict balancing before patches are loaded.

Fixed FunPack Hybrid Euler 2S Sampler restart timing so restart_trigger_pct is respected across the full sigma schedule instead of being clamped to the Euler-to-2S quality transition.

Improved FunPack LoRA Loader rerun performance by caching recently used raw LoRA files, model-mapped LoRA patches, and per-block fingerprint analysis.

Reworked FunPack Video Refiner V2 ratings around explicit missing-axis signals: Perfect, single missing axes, paired missing axes, and Awful.

Removed the Refiner V2 mode input. V2 now accepts whatever connected CLIP the workflow provides and stores state in a CLIP-owned namespace.

Renamed visible Refiner and LoRA intent from concept to action. Old Missing concept ratings and old concept LoRA rows are still accepted as compatibility aliases, but V2 stores and displays action.

Updated I'm Feeling Lucky in Refiner V2 so Lucky only composes prompt text when enabled. When disabled, it may train memory from rated runs but does not compose or alter output.

Fixed

Fixed I'm Feeling Lucky token-bank learning for changing prompt/conditioning workflows by falling back to prompt-order token placement when exact tokenizer position matching cannot find enough words.

Fixed I'm Feeling Lucky rating attribution for changing prompts so ratings update the previous prompt's learned tokens while the current prompt seeds new neutral discovery tokens.

Updated I'm Feeling Lucky filtering to learn poor adjacent token pairs instead of refusing individual tokens outright.

Updated I'm Feeling Lucky with uncapped token, pair, and context memory so it can learn which concepts belong together and call strong missing neighbors when prompt anchors are present.

Fixed I'm Feeling Lucky composition order so the current generation uses already-learned memory first, then seeds current prompt tokens for future runs.

Fixed I'm Feeling Lucky memory-first output so vague or empty incoming conditioning can use the longest compatible learned conditioning canvas instead of being limited to the current prompt's shape/content.

Added an optional clip input to FunPack Video Refiner so I'm Feeling Lucky can compose a learned prompt, re-encode it through the connected CLIP/Gemma text encoder, and refine from that freshly tokenized conditioning.

Improved I'm Feeling Lucky runtime by selecting learned conditioning canvases from saved tensor metadata before decoding, capping CLIP/Gemma Lucky prompts to a practical per-run concept count, and decoding only the token vectors selected for the current generation.

Reduced redundant I'm Feeling Lucky work by keeping Lucky runs in one stable memory history, skipping normal prompt-variant conditioning scans while Lucky is active, validating large Lucky memories once per loaded session, and updating context relationships locally instead of writing all-to-all token graphs every run.

Fixed I'm Feeling Lucky CLIP/Gemma re-encode crashes when the encoded Lucky prompt has a different sequence length than the incoming conditioning by resizing the refinement delta before applying it.

Updated I'm Feeling Lucky CLIP/Gemma prompt composition to preserve learned comma/semicolon-separated concept phrases instead of emitting loose word lists when phrase memory is available.

Added Lucky phrase placement memory so learned prompt phrases remember their rated order positions and CLIP/Gemma Lucky prompts can reassemble phrases into a more coherent prompt order instead of sentence salad.

Fixed I'm Feeling Lucky bootstrap learning so sessions that start with Lucky enabled now create a real discovery history entry, seed prompt tokens/phrases, and can learn from ratings without first running the classic refinement loop.

Updated Lucky memory so normal non-Lucky runs still seed reusable token, phrase, context, and placement memory for later Lucky runs.

Updated all missing-axis ratings so Missing details, Missing concept, Missing quality, and paired missing ratings now mark prompt tokens as wanted-but-underrepresented instead of weak neutral feedback; repeated missing feedback reserves Lucky composition room for those tokens and their compatible neighbours.

Fixed Lucky diagnostics so the collapsed Lucky memory stream reports real Lucky update counts and learned memory size instead of implying the session is still prompt 1 out of 1.

Removed

Removed the old public FunPack Video Refiner node, the FunPackGemmaEmbeddingRefiner compatibility alias, and FunPack Save Refinement Latent from the registered node list.

Removed sigma refinement, latent refinement, manual scheduler controls, and feedback-question inputs from the active Refiner workflow. These systems are not part of Refiner V2.

[2.1.3] - 2026-04-24

Changed

FunPack Apply LoRA Weights now has more user-friendly, compact UI.

FunPack Video Refiner now has updated logic to work more stable when provided different prompts and conditioning with each new generation.

[2.1.1] - 2026-04-24

Added

Added -Just forget it- as a Video Refiner rating. Use it when a generation failed for reasons that should not be learned from, such as a broken reference, bad seed, or workflow mistake.

Added category feedback questions for prompt phrases that the refiner cannot confidently classify. The answer scale is general, concept, style, quality, character, and details.

Added a CLIP Vision output combiner node for workflows that need one combined CLIP Vision output from multiple inputs.

Changed

Updated the Video Refiner rating categories so feedback can separate missing details, missing concept, missing quality, and fully failed output instead of treating all bad results the same way.

Reduced repeated category feedback prompts after the user has already answered enough about the same concept.

Refreshed README and refiner docs for 2.1.1.

Fixed

Fixed LoRA weight row restore order when workflows are loaded.

[2.1.0] - 2026-04-23

Added

Added FunPack Apply LoRA Weights and FunPack LoRA Loader, a prompt-exact LoRA weight workflow designed to work with FunPack Video Refiner.

Added FunPack Save Refinement Latent, which stores latent tensor bundles by refinement key for optional latent refinement in FunPack Video Refiner.

Added hidden LTX per-block LoRA redistribution for supported ltx2 model stacks. The UI still exposes normal LoRA weights, while the loader derives per-block strengths from the LoRA patch magnitudes when the model and LoRA layout support it.

The new workflow uses base LoRA weights on the first run for a prompt, then lets the refiner save prompt-specific suggested LoRA weights into its existing JSON state for later runs.

Changed

Renamed the visible refiner title from FunPack Gemma Embedding Refiner to FunPack Video Refiner. The old node key is still available as a compatibility alias.

Split the old single funpack.py implementation into focused modules:

• conditioning.py
• samplers.py
• image_processing.py
• model_management.py

funpack.py remains as a compatibility re-export for older imports.

Updated FunPack Video Refiner so it can accept a FunPack LoRA stack and save next-run model LoRA weight suggestions based on prompt concepts, LoRA type hints, and user ratings.

Updated FunPack Video Refiner with optional latent input/output refinement. If no matching saved latent exists and both latent input and output are connected, the input latent is saved as the first reference and passed through unchanged.

Updated prompt analysis so quoted speech and backslash-wrapped phrases can be protected as whole prompt units.

Documentation

Documented unintended and edge-case usage for the new refiner workflow, including disconnected latent paths, saved-latent-only runs, wrong LTX audio/AV latent connections, exact-prompt LoRA lookup behavior, base-weight mismatch behavior, zero-weight LoRA skipping, and unsupported per-block fallback behavior.

[1.3.3] - 2026-04-22

Changed

Expanded /docs so every node in funpack.py now has matching documentation, and refreshed the existing node docs to match the current inputs and outputs.

[1.3.2] - 2026-04-19

Changed

Changed the core logic of Self-Refiner. Removed obsolete nodes.

[1.3.0 & 1.3.1] - 2026-04-18

Changed

Added new nodes - User Rating and Gemma Self-Refinement for LTX2.3 video workflows.

Fixed

Device type mismatch in new nodes.

[1.2.3] - 2026-01-30

Fixed

Fixed Transformers library error when running Prompt Enhancer and Story Writer nodes.

[1.2.2] - 2026-01-26

Changed

Changed the logic of processing sequences in Story Writer node. Now doesn't append full instructions and previous context to previous messages with each loop iteration, now fully replaces messages with a system prompt and sequence history without appending.

[1.2.1] - 2026-01-24

Added

Added experimental LoRA recommendation feature and Sanity Check features to Story Writer node.

[1.2.0] - 2026-01-23

Added

Added new Story Writer node, based on existing Prompt Enhancer. It generates up to 5 prompts one after another, based on either user's prompt directly, or on the story generated from the user's prompt.

[1.1.0] - 2026-01-02

Added

Added Creative Template and Lorebook Enhancer nodes. The Creative Template is a wildcard-based node that replaces given keywords in the template with ones provided by user. Lorebook Enhancer is a node that takes SillyTavern format .json lorebooks and enhances your prompt by adding required knowledge.

[1.0.0] - 2026-01-01

Initial release on Comfy Registry.