nodes/olivv-cs-ComfyUI-FunPack
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
2,225 Commits 3 Branches 11 Tags
d5ec8280263ae1c435a6cc4239a343c73bb41b80
Commit Graph
56 Commits
Author SHA1 Message Date
Claude 129cc10d0f fix: video and audio references go to R2V's own video/audio inputs; aspect ratio source accepts a video 
New Load Media (video) and (audio) nodes; picture loader keeps first frame for the aspect source only.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CHWxqHtxwWrq4kEyFTaHds
 2026-10-21 02:01:01 +00:00
Claude 919097c732 LoRA training: our own loop for H3, and a Log that survives a crash 
The H3 run trains with a loop of our own (LoRA on the attention and feed-forward projections of the
50 blocks, per-block checkpoint, uniform noise levels, musubi-tuner's approach) while ComfyUI carries
the batches. Every Log record is flushed to disk as it is said and reloaded on start; the run
writes each stage and step to it. ComfyUI's own offloading toggle is ignored by the new loop.
UNTESTED on the real weights.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 19:57:06 +00:00
Claude a13f037f72 H3 training: modulate blocks out of place so the backward pass stops keeping every block's recompute alive 
Measured on a tiny H3: backward growth was ~75 MB per block, now flat (61-88 MB at 2, 8 and 16 blocks).
The in-place mul_/addcmul_ on slices of the stream in the block helpers were the cause.
UNTESTED on the real weights.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 19:47:34 +00:00
Claude d0d572c4ec feat: block swap option for LoRA training and a per-block memory log in the backward pass 
Block swap keeps the frozen weights of the last N transformer blocks in system
memory and lets ComfyUI's weight casting fetch them one layer at a time. The idea
and the block count as the one knob follow musubi-tuner's block swap (kohya-ss,
Apache License 2.0); this is an independent implementation.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 19:30:50 +00:00
Claude a070ce9400 Training reads every clip for as many frames as the shortest gives, and allows up to 1025 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 18:30:50 +00:00
Claude bb9ec64a46 H3 training checkpoints one block at a time and the text refiner on its own 
Depth 1 made the whole model a single checkpoint, so nothing was saved.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 18:14:55 +00:00
Claude ab100f85a4 LoRA training add-on: sets, tagger, trainer node on ComfyUI's Train LoRA, H3 support 
Backend only (window follows). Switched on in Settings > Modules. H3 repairs proven on a tiny H3 on CPU, UNTESTED on real weights.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
 2026-10-17 15:05:57 +00:00
digital-garbageandClaude Sonnet 5.5 1565ac3982 fix: skipped slots offer no Remove/Swap, stand-in load clears skipped, null pass value refused; route-level and precedence tests 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-16 01:28:22 +03:00
digital-garbageandClaude Sonnet 5.5 803ea68f21 fix: optional nodes are declared per node class by the model modules (saved pipelines and swapped nodes covered); skip chain cannot crash; skipped controls hidden in the project tab; skipped card says so (attacker round 26) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-16 01:02:59 +03:00
digital-garbageandClaude Sonnet 5.5 3beb13f6c5 feat: optional preset slots - a pipeline runs without Enhance, Shot camera, Prompt markup, Negative erase, Cut opening or Identity transfer when their module did not load, says so beside Generate, and hides their controls 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-16 00:41:38 +03:00
digital-garbageandClaude Sonnet 5.5 08550e3569 fix: H3 second-pass preset held back; unsupported .gguf text encoders are refused instead of loading as the wrong model; Start shots hint is true for H3; LTX module titled UNTESTED; docs corrected (failure behaviour, Taste key location, install caution, gguf, second pass) (attacker round 5) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-15 18:10:57 +03:00
digital-garbageandClaude Sonnet 5.5 43e63ebf74 docs: v5 CHANGELOG, how-to-start guide, release README (Intent and Thank you kept verbatim), codename note; LTX declared UNTESTED loudly (preset titles, every run's log, docs) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-15 17:53:00 +03:00
digital-garbageandClaude Sonnet 5.5 648ea94c4a fix: text-only runs do not hand the grey stand-in to the enhancer; the encoder pick is text so a vanished file cannot stop a run with Enhance off; truer retain_first, montage and upscaler hints (attacker round) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-15 16:59:48 +03:00
digital-garbageandClaude Sonnet 5.5 1fc5d99dcc feat: MiniMax H3 Reference to Video works text-only; with no picture a plain stand-in of the Width/Height shape feeds the size chain (found on a real GPU) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-15 15:03:52 +03:00
digital-garbage 23b46559cc fix: hints and labels now say what the control does (menus, inspector, timeline, takes, ideas, enhancer, context windows, montage, modes); H3 project negative prompt reaches the negative slot 2026-10-15 12:55:40 +03:00
digital-garbageandClaude Sonnet 5.5 d59a9ecba4 fix: a loaded project file is saved cleared; the shot camera node always gets the project id; infusion skips a word the shot already says 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-13 07:11:54 +03:00
digital-garbageandClaude Sonnet 5.5 090f8cfbb8 feat: focus infusion draws only on words the current project's own prompts keep returning to 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-13 06:29:41 +03:00
digital-garbageandClaude Opus 5.5 99f00f2950 feat: FunPack Save Video takes frames, sound and fps directly; Create Video dropped from the presets 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-07 16:08:28 +03:00
digital-garbageandClaude Opus 5.5 82d18494ad feat: FunPack Save Video: H.265 on NVENC (hvc1, plays in Mac Chrome/Safari), H.264 option; no NVENC = H.264 on every CPU core (core SaveVideo used one); presets use it 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-07 15:43:18 +03:00
digital-garbageandClaude Opus 5.5 05fde80caf feat: H3 decode logs where its time goes: picture vs sound seconds, VAE precision, and whether the VAE was fully on the GPU 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-07 15:26:31 +03:00
digital-garbageandClaude Opus 5.5 109ea2fb4f revert: H3 decode in one piece showed a tile grid (the decoder is trained on 256px tiles); back to ComfyUI's own tiling, tooltip says so 
Reverts b015ce3f.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-07 15:01:07 +03:00
digital-garbageandClaude Opus 5.5 b015ce3f87 fix: H3 decode with tile size 0 decodes each frame in one piece (core tiled at 256 regardless); out of memory steps down to 512, then 256 tiles 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-07 14:51:11 +03:00
digital-garbageandClaude Opus 5.5 ef84df4286 fix: views redraw when a project's pipeline goes in; project switches wait for edits in flight and load one at a time; modifier settings hide when the pipeline cannot install them; refusals name the setting the person sees; older SD1.x files are recognised; stale load retries are cancelled 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-05 09:23:18 +03:00
digital-garbageandClaude Opus 5.5 6fa6c4c6ee fix: a preset, import or wizard pipeline reopens exactly as saved (roles too); Taste key hides when no learning feature is on; a broken node schema no longer takes the module list down; a failed model probe is retried 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-05 04:40:32 +03:00
digital-garbageandClaude Opus 5.5 927d8d7358 feat: Stable Diffusion 1.5 model module, so the whole pipeline runs on a Mac (picture held for the clip's length) 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-10-05 04:19:20 +03:00
digital-garbageandClaude Sonnet 5.5 054f18a525 fix: Generate reads the live screen, not a save; autosave delay is a setting that waits for typing; H3 preset always carries a LoRA loader (None = untouched) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 15:51:40 +03:00
digital-garbageandClaude Sonnet 5.5 dce64c1813 feat: shot camera (cut times, views, camera moves for [Shot N]) with rating-taught memory 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 14:07:03 +03:00
digital-garbageandClaude Sonnet 5.5 3f8607b343 feat(h3): Negative Erase node — the negative prompt takes its direction out of the positive at CFG 1 (off at strength 0) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 12:05:10 +03:00
digital-garbageandClaude Sonnet 5.5 1343486f87 feat(ltx): 'Anchor as guide' pipeline preset (stock LTXVAddGuide at frame 0, empty latent, guides cropped before decode) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 12:01:55 +03:00
digital-garbageandClaude Sonnet 5.5 69c678e97e fix(ui): a run already finished at reload is recorded at once instead of locking the page; recovered length uses the shared-frame rule; H3 second pass gets a fresh seed too 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 20:02:46 +03:00
digital-garbageandClaude Sonnet 5.5 e0a23b93b5 fix(ui): fresh seed per run (generation.seed role), a unit is made at its clips' length, all-excluded units skip (not just the root), monitor stops at a clip's own length, inspector follows the pipeline, settings capped at the file's limit 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 19:46:14 +03:00
digital-garbageandClaude Sonnet 5.5 d14bc51c2d fix(ui): left-edge trim direction, one source for length/fps (roles 'drive' it), whole-number settings only, stop/queue/no-prompt/adopt messages, ghost remove, cut halves read-only prompt, preview start+end; font test sees app/ui fonts 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 19:32:36 +03:00
digital-garbageandClaude Sonnet 5.5 a51faa729b feat: LTX modules — video->audio coupling, context windows, identity transfer (Best-FaceID); TinyLTX real-model fixture; ltx_av trait 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 10:48:36 +03:00
digital-garbageandClaude Sonnet 5.5 47752491db fix: LTX — tiled decode asked the way comfy asks it, picture found for ALG/DynaShift/guidance (shared av_video_stream), odd sizes said 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 10:30:37 +03:00
digital-garbageandClaude Sonnet 5.5 6bfeb63c4c feat: LTX 2.x model module — recognised from the checkpoint header, joint video+audio latent sized from the audio VAE, decode, text/image-to-video pipelines 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-03 10:24:43 +03:00
digital-garbageandClaude Sonnet 5.5 fd236a6e97 fix: phrase probe re-installs when its switch flips (cache key), sits innermost so other features' extra forwards are not measured, panel keeps peaks; enhancer reads Qwen-style tokenizers, chat comments key on the typed prompt; scenes get eff_frames/eff_fps 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-02 15:35:37 +03:00
digital-garbageandClaude Sonnet 5.5 8be3770080 feat: H3 tiled and X2 Detail VAE decode, phrase probe with Settings panel 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-02 15:17:03 +03:00
digital-garbageandClaude Sonnet 5.5 b29fda7a0e feat: prompt enhancer (node module, H3 preset slot, Composer Enhance + Chat tabs, readout route) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-02 15:05:22 +03:00
digital-garbageandClaude Opus 5.5 0bb7e3a01a fix: H3 picture found in the packed model output; latent steering was off on every real run 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 19:01:13 +03:00
digital-garbageandClaude Opus 5.5 f99944e5e4 fix: DynaShift and taste guidance read the step from ComfyUI's positional call; they never steered 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 18:53:51 +03:00
digital-garbageandClaude Opus 5.5 b831372879 feat: DynaShift steers late steps away from disliked clips, banked on the taste key 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 18:34:16 +03:00
digital-garbageandClaude Opus 5.5 53cbc29948 feat: taste steering + taste attention learn from ratings; a failing wrapper no longer re-runs sampling 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 18:31:10 +03:00
digital-garbageandClaude Opus 5.5 9d028ccc92 feat(v5): prompt markup for H3 -- (word:1.5), [phrase@t0-t1], [a|b] 
Prompt Markup strips it before the encoder; Apply Prompt Markup puts weights
and timed windows on attention and blends on the conditioning. Where the words
sit is the H3 module's (text_tokenizer, prompt_rows, video_time_rows). Wired
into both H3 presets.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 18:11:51 +03:00
digital-garbageandClaude Opus 5.5 6838e3b997 feat(v5): cut opening frames -- keep the anchor's transfer, drop the still 
Cuts decoded frames and the same time off the sound. In both H3 presets,
0 by default.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 17:56:15 +03:00
digital-garbageandClaude Opus 5.5 5787190b3f feat(v5): H3 second pass as graph -- upscaler loader, latent resample, typed pass-2 schedule 
H3's latent upscaler loads through the H3 module (ComfyUI's loader has no
branch for it); an upscale resizes H3's anchor pin so pass 2 stays pinned.
FunPack Sampler takes an optional wired schedule, run exactly as written.
New preset: H3 Reference to Video + second pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 17:49:39 +03:00
digital-garbageandClaude Opus 5.5 56490ec769 feat(v5): ALG loosens H3's pinned opening keyframe 
The H3 module offers anchor_pin, so ALG blurs the pin's latent without
learning H3's payload layout.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 17:37:49 +03:00
digital-garbageandClaude Opus 5.5 c3ad0dfd79 feat(v5): randomizer and block repeat modules for H3, on shared chainable block hooks 
Block repeat's video-only mode now really keeps audio single-pass: H3 blocks
write in place, so v4 restored audio from an already-repeated stream.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 17:29:42 +03:00
digital-garbageandClaude Opus 5.5 ec6e9893b4 feat: hide settings the picked model cannot use, as soon as the file is picked 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
 2026-09-27 17:20:59 +03:00
digital-garbageandClaude Sonnet 5 776529e0d8 feat: assign category to every remaining module 
The port plan calls for every module under modules/ to declare a
CATEGORY so Settings/Engine panels can group by it. Only alg, sharpen,
appearance, and full_control had one; the other 12 (loaders, empty
latent, media load, minimax_h3, output decode, sampling modifiers/
sampler, system pipeline) are plain infrastructure with no user-facing
settings dial, so they land in "system" except output_decode (post,
matching its STAGE) and the two sampling modules (sampling). Verified
against core/tests (548 passed) -- category validation and the
registry's field-mapping catch a typo here, per the earlier fixture-gap
fix.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
 2026-09-15 09:46:58 +03:00
digital-garbage 4744d30087 fix: H3 preset combo values rejected by real node schemas 
keep_proportion/extra_padding/invert_crop (ImageTransformKJ) and format
(SaveVideo) were wrapped in a redundant {"field": "value"} dict instead of
the plain string ComfyUI's own DynamicCombo expansion expects at the wire
level -- every one of them silently failed graph validation and fell back
to the node's own default, undermining the "total_pixels" resolution-fit
behavior this preset exists for. Found by loading the preset in a real
browser against real node schemas; the dev server never had comfy_extras
or KJNodes registered before now, so this was invisible locally until
tools/devserver.py also gained real node registration (init_extra_nodes +
KJNodes' load_custom_node, matching what the pytest `registered` fixture
already did).

The two pipeline_preset tests that were supposed to catch this had too
broad a filter -- it excluded ANY "is not one of" message as an assumed
"combo not chosen yet" state, which in this codebase actually shows up as
"needs X and nothing fills it" instead. Tightened to exempt only the three
loader-file combos genuinely left unset on a fresh install.

Verified live: preset loads with zero combo-validation problems, only the
three expected unset-loader-file messages remain, and Generate correctly
refuses with an honest reason rather than a false success.
 2026-09-11 13:32:52 +03:00