Docs and changelog for 2.5.0 release
- CHANGELOG: add new Changed and Fixed entries for negative conditioning removal, advisor system prompt rewrite, token limit increases, Only prompt double-call fix, and encoded_prompts bug fix. - README: update 2.5.0 section to cover all of the above. - docs/FunPackVideoRefinerV2.md: remove negative_prompt input and modified_negative output; update advisor_mode description with new structured input format and token budgets; update encoded_prompts output description with all four possible sections. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
digital-garbage
co-authored by
Claude Sonnet 4.6
Claude Sonnet 4.6
parent
bb25450963
commit
78aae515b4
+43
-5
@@ -1,16 +1,54 @@
|
||||
# Changelog
|
||||
|
||||
## Unreleased
|
||||
## [2.5.0] - 2026-05-15
|
||||
|
||||
### Added
|
||||
|
||||
Added an optional CLIP text-generation advisor to `FunPack Video Refiner V2`. The advisor uses an explicit Refiner system prompt, can run Gemma thinking mode, passes source images to compatible CLIP tokenizers, sees the prompt that caused the current rating, and can either report diagnostics or apply a validated prompt repair before encoding.
|
||||
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 an optional `advisor_clip` input to `FunPack Video Refiner V2` so the advisor can use a separate generative CLIP/Gemma model while the main `clip` continues handling normal prompt encoding.
|
||||
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 a single `encoded_prompts` string output to `FunPack Video Refiner V2` with `Positive prompt:` and `Negative prompt:` sections, so workflows can inspect the exact prompt text used for conditioning.
|
||||
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.
|
||||
|
||||
Updated the Refiner V2 advisor so `Repair prompt` can advise the negative prompt as well as the positive prompt when a rating indicates wrong or strongly bad content.
|
||||
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
|
||||
|
||||
|
||||
@@ -2,6 +2,22 @@
|
||||
|
||||
A set of ComfyUI nodes for experimenting with video generation workflows based on WAN, HunyuanVideo, LTX, and similar models.
|
||||
|
||||
## Updates in 2.5.0
|
||||
|
||||
Added a two-pass CLIP text-generation advisor to `FunPack Video Refiner V2`. Pass 1 analyses what specifically needs to change in the suggested prompt. Pass 2 applies those findings. The advisor uses a structured input format — `ORIGINAL_USER_INTENT`, `LAST_PROMPT`, `RATING`, and `OPTIONAL_NOTE` — so the model knows exactly what failed and what to fix. Token budget: 1200 tokens for analysis, 1600 for repair.
|
||||
|
||||
Added `feedback_prompt` input. When connected, the user's natural-language description of what was wrong is placed first in both advisor passes and the system follows it exactly, overriding axis-based repair logic and validation guards.
|
||||
|
||||
Added `Prompt only` execution mode: all prompt shaping runs normally but conditioning vectors pass through unchanged. Added `prompt_repair` boolean to disable the rule-based phrase injection from memory when not enough context has been built yet.
|
||||
|
||||
Added `encoded_prompts` STRING output. When the advisor ran it shows up to four sections: `Positive prompt` (what was encoded), `Advisor suggestion (applied/rejected)` (the advisor's generated prompt), `Advisor analysis` (diagnostic 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 decays toward this value as sigma approaches the quality phase boundary. The early phase now uses order-2 denoised extrapolation (Adams-Bashforth 2-step) at zero extra model-call cost, and the quality phase uses progressive correction blending to reduce the number of expensive 2S evaluations.
|
||||
|
||||
Removed `negative_prompt` input and `modified_negative` conditioning output from Refiner V2. Negative conditioning has no effect at CFG=1.0 and was adding a redundant generation call to every advisor run.
|
||||
|
||||
Fixed greedy decoding in the advisor: `do_sample` was `False`, causing the model to always produce its highest-probability default output. Fixed `Only prompt` mode running two generation calls per invocation (positive + the now-removed negative advisor). Fixed `encoded_prompts` always showing only the positive prompt regardless of advisor activity. Also fixed four logic bugs: intent example lookup reading a nonexistent field, streak updates contaminating conditioning strength in Prompt only mode, unfiltered global phrase memory injecting unrelated vocabulary, and the first-run rating label being forwarded when no previous output existed.
|
||||
|
||||
## Updates in 2.4.2
|
||||
|
||||
Added Refiner V2 `Learning` mode. It observes prompts, conditioning, ratings, phrase memory, and diagnostics while passing positive and negative prompt conditioning through unchanged.
|
||||
@@ -26,51 +42,9 @@ Added advisory image/CLIP Vision context and repaired negative conditioning to `
|
||||
|
||||
Added an opt-in experimental early velocity bias mode to `FunPack Hybrid Euler 2S Sampler` for capturing/applying averaged early denoise directions around normalized sigma 0.9 and 0.8.
|
||||
|
||||
## Updates in 2.3.3
|
||||
## Updates in 2.3.x
|
||||
|
||||
Refiner V2 now supports the restored pre-encoded conditioning workflow: connect `positive_conditioning` without `CLIP` and it accepts the finished Gemma3/LTX2 conditioning while loading only the Gemma3 tokenizer. If `CLIP` is connected, V2 still encodes the prompt itself and ignores the fallback conditioning path.
|
||||
|
||||
## Updates in 2.3.2
|
||||
|
||||
Refiner V2 now learns original-intent alignment when `user_intent_prompt` stays the same but an enhancer gives different `positive_prompt` variants. Ratings teach it which intent-enhance pairs represented the original request, which original phrases were missing, and which enhancer-only additions were rejected.
|
||||
|
||||
Learned original-intent omissions can now be restored on later runs, while repeatedly rejected enhancer-only full words and adjacent word pairs can be omitted before encoding.
|
||||
|
||||
## Updates in 2.3.1
|
||||
|
||||
Refiner V2 Prompt Repair now keeps all missing/wrong ratings tied to the current prompt or explicit user intent. Learned favorite actions, details, quality cues, camera moves, and styles are not repaired into unrelated requests just because they scored well before.
|
||||
|
||||
Prompt Repair now treats the same word in different neighbor contexts as separate evidence, so a liked phrase does not automatically transfer to a different request that happens to share one word.
|
||||
|
||||
When the optional raw user intent is vague, such as `Figure it out`, Refiner V2 now treats the enhanced `positive_prompt` as the stronger repair anchor.
|
||||
|
||||
## Updates in 2.3.0
|
||||
|
||||
Refiner V2 now blocks learned appearance, character, subject, and background concepts from Prompt Repair, Lucky auto-injection, and legacy Void memory unless the current prompt explicitly asks for them.
|
||||
|
||||
Added `Wrong appearance` for outputs polluted by remembered clothing or character traits. This suppresses the responsible appearance memory without penalizing unrelated action, camera, detail, or quality learning.
|
||||
|
||||
Added `FunPack Refinement Key Loader` for selecting, creating, importing, and exporting Refiner V2 keys. The loader can feed both Refiner V2 and `FunPack Apply LoRA Weights`.
|
||||
|
||||
Added a paste-friendly [`Refiner V2 quick guide`](docs/FunPackVideoRefinerV2QuickGuide.md) for new users.
|
||||
|
||||
## Updates in 2.2.1
|
||||
|
||||
Fixed Refiner V2 prompt phrase categorization so background, appearance, quality, camera, and action phrases stay aligned before the node updates Lucky memory or LoRA suggestions.
|
||||
|
||||
## Updates in 2.2.0
|
||||
|
||||
Added `FunPack Video Refiner V2`, a simpler prompt-owned refiner that takes `positive_prompt` and a connected `CLIP`, encodes the prompt inside the node, learns from ratings, and returns refined positive conditioning plus diagnostics.
|
||||
|
||||
Removed Refiner V2's `mode` input. The node now accepts whatever connected `CLIP` the workflow provides and stores V2 state in a CLIP-owned namespace.
|
||||
|
||||
Refiner V2 removes the old sigma refinement, latent refinement, manual scheduler controls, and feedback question workflow. It now adapts internally: consistently good ratings make updates gentler, while consistently bad ratings make updates stronger.
|
||||
|
||||
Renamed visible Refiner and LoRA intent from `concept` to `action`. Internally, `action` means action plus motion, so movement verbs, physical motion, subject motion, and camera motion are treated together. Old `Missing concept` ratings and old `concept` LoRA rows are accepted as aliases.
|
||||
|
||||
Updated `I'm Feeling Lucky` in Refiner V2 so Lucky is only a prompt composer. When Lucky is off, it can still train memory from rated runs, but it does not compose or alter the output.
|
||||
|
||||
Removed the old public `FunPack Video Refiner`, its compatibility alias, and `FunPack Save Refinement Latent` from the node list. Use `FunPack Video Refiner V2` for new refinement workflows.
|
||||
Refiner V2 now supports pre-encoded conditioning workflows, original-intent alignment memory, improved prompt repair scoping, and a `Wrong appearance` rating for outputs polluted by remembered appearance concepts. Added `FunPack Refinement Key Loader`.
|
||||
|
||||
## Dev Branch
|
||||
|
||||
|
||||
@@ -1,16 +1,12 @@
|
||||
# FunPack Video Refiner V2
|
||||
|
||||
Video Refiner V2 is a simplified prompt-owned refiner. It takes prompt text and a connected `CLIP`, encodes the prompt inside the node, learns from ratings, and returns refined positive conditioning.
|
||||
Video Refiner V2 is a prompt-owned refiner. It takes prompt text and a connected `CLIP`, encodes the prompt inside the node, learns from ratings, and returns refined positive conditioning.
|
||||
|
||||
V2 replaces the old `FunPack Video Refiner` public node in version `2.2.0`. It uses separate fresh state and does not migrate old histories.
|
||||
V2 replaces the old `FunPack Video Refiner` public node. It uses separate fresh state and does not migrate old histories.
|
||||
|
||||
## Inputs
|
||||
|
||||
**positive_prompt**: Prompt text to encode and refine.
|
||||
|
||||
**clip**: Connected text encoder. V2 uses this CLIP for prompt encoding, phrase/category similarity checks, and optional advisor text generation when the connected CLIP supports `generate`/`decode`.
|
||||
|
||||
**advisor_clip**: Optional separate CLIP/Gemma text generator for the Advisor. When connected, Advisor uses this model for text generation while the main `clip` still owns prompt encoding and similarity checks. When disconnected, Advisor falls back to `clip`. If neither selected CLIP exposes text generation, Advisor skips and V2 continues with the usual workflow.
|
||||
**positive_prompt**: Prompt text to encode and refine. When an LLM enhancer is in the workflow this is the enhancer output. V2 treats it as the "suggested prompt" when building advisor context.
|
||||
|
||||
**rating**: Feedback for the previous V2 generation:
|
||||
|
||||
@@ -30,90 +26,101 @@ V2 replaces the old `FunPack Video Refiner` public node in version `2.2.0`. It u
|
||||
|
||||
Older `Missing concept` labels are accepted as aliases for `Missing action`.
|
||||
|
||||
Use `Wrong appearance` when the video is polluted by remembered clothing, character traits, subject identity, or background concepts that were not requested for the current image. This suppresses those concepts from future auto-injection without deleting them for explicit prompts.
|
||||
Use `Wrong appearance` when the video is polluted by remembered clothing, character traits, subject identity, or background concepts that were not requested for the current image.
|
||||
|
||||
**refinement_key**: V2 session key. V2 stores separate `refine_v2` state under this key.
|
||||
**mode**: Execution mode:
|
||||
|
||||
**refinement_key_input**: Optional linked key input, usually from `FunPack Refinement Key Loader`. When connected, it overrides the text widget.
|
||||
- `Refine`: applies learned prompt shaping and conditioning vector adaptation. Default.
|
||||
- `Prompt only`: applies all prompt shaping but passes conditioning vectors through unchanged. Learning still runs. Use when you want to pause conditioning adaptation while prompt refinement continues.
|
||||
- `Learning`: observes ratings and updates all memory, but passes both prompt and conditioning through unchanged.
|
||||
|
||||
**advisor_mode**: CLIP text-generation advisor:
|
||||
|
||||
- `Off`: no advisor call.
|
||||
- `Only diagnostics`: runs the analysis pass only, stores the finding in feedback history as an `Advisor note:` entry, and returns the prompt unchanged. Use this to preview what the advisor would say before committing to `Full`.
|
||||
- `Only prompt`: runs the repair pass only (no analysis). Changes the prompt silently; outputs a plain prompt string with no labels.
|
||||
- `Full`: runs both passes. Pass 1 identifies what specifically needs to change. Pass 2 applies those findings. The advisor's suggestion, analysis text, and pre-repair prompt are all visible in `encoded_prompts`.
|
||||
|
||||
The advisor receives four explicit inputs: `ORIGINAL_USER_INTENT` (the user's intent prompt), `LAST_PROMPT` (the previous encoded prompt), `RATING` (the current rating label), and `OPTIONAL_NOTE` (feedback prompt, memory suggestions, history, and analysis combined). Token budget: 1200 for the analysis pass, 1600 for the repair pass.
|
||||
|
||||
**clip**: Text encoder. V2 uses this for prompt encoding, phrase and category similarity checks, and advisor text generation when the model exposes `generate`/`decode`.
|
||||
|
||||
**advisor_clip**: Optional separate generative CLIP/Gemma model for the advisor. When connected, the advisor uses this model for both analysis and repair passes while `clip` continues handling encoding and similarity. When disconnected, the advisor falls back to `clip`.
|
||||
|
||||
**advisor_thinking**: Enables thinking mode for compatible Gemma CLIP text generators.
|
||||
|
||||
**feedback_prompt**: Optional natural-language description of what was wrong with the previous output, e.g. `he was supposed to hold her hand not her head`. When connected it is placed first in both advisor passes and the system follows it exactly, bypassing axis-based repair logic, intent distance checks, and protected category rules.
|
||||
|
||||
**prompt_repair**: Enables or disables the rule-based phrase injection from phrase memory for missing axes (default on). Turn off early in a session before enough context has accumulated, or when memory suggestions are disrupting the generation.
|
||||
|
||||
**refinement_key**: V2 session key. All memory and history is stored under this key.
|
||||
|
||||
**refinement_key_input**: Optional linked key input, usually from `FunPack Refinement Key Loader`. When connected, overrides the text widget.
|
||||
|
||||
**seed**: Optional seed for deterministic exploration paths.
|
||||
|
||||
**reset_session**: Clears the V2 session state for the selected key.
|
||||
**reset_session**: Clears all V2 session state for the selected key. Scene Builder memory is preserved.
|
||||
|
||||
**lora_stack**: Optional stack from `FunPack Apply LoRA Weights` / `FunPack LoRA Loader`. V2 writes prompt-specific suggested model weights.
|
||||
|
||||
**clip_vision_output**: Optional CLIP Vision output for the original/source image. V2 stores tensor summaries as advisory context and diagnostics only; it does not blend this into positive conditioning.
|
||||
**clip_vision_output**: Optional CLIP Vision output for the source image. Stored as advisory context and diagnostics only; not blended into positive conditioning.
|
||||
|
||||
**source_image**: Optional source image/frame batch. V2 stores width, height, aspect ratio bucket, and a small fingerprint so it can report whether the input image changed between runs.
|
||||
**source_image**: Optional source image or frame batch. V2 stores width, height, aspect bucket, and a fingerprint to report when the input image changed between runs. Also passed to vision-capable advisor models.
|
||||
|
||||
**negative_prompt**: Optional negative prompt. V2 can persistently add poorly rated or wrong-context tags to this prompt, then encode the repaired prompt as negative conditioning when `CLIP` is connected.
|
||||
**user_intent_prompt**: Optional raw user request before any LLM enhancement. V2 compares it with `positive_prompt`, stores intent-enhance pairs, and can restore omitted original phrases or suppress repeatedly rejected enhancer-only additions.
|
||||
|
||||
**user_intent_prompt**: Optional raw/original request. When connected, V2 compares it with the enhanced `positive_prompt`, stores intent-enhance pairs, remembers all provided full-word and adjacent word-pair tokens for that intent, and can omit enhancer-only tokens that repeatedly led to bad ratings.
|
||||
**im_feeling_lucky**: Composes a learned prompt from rated phrase memory and encodes it through the connected CLIP.
|
||||
|
||||
**im_feeling_lucky**: When enabled, V2 composes a learned prompt from rated phrase memory and encodes it through the connected CLIP. When disabled, V2 may still store Lucky memory from rated runs, but it does not compose or apply a Lucky prompt.
|
||||
## Feedback History
|
||||
|
||||
**advisor_mode**: Optional CLIP text-generation advisor:
|
||||
V2 maintains a rolling history of the last ten user-provided `feedback_prompt` entries in session state. Each entry is labelled with its rating, for example:
|
||||
|
||||
- `Off`: no advisor call.
|
||||
- `Diagnostics`: asks the connected generative CLIP for a short diagnosis, but never changes the encoded prompt.
|
||||
- `Repair prompt`: asks for a repaired positive prompt and applies it only after validation against the current request, original intent, refusal filter, and protected appearance/subject/environment rules.
|
||||
```
|
||||
1. Missing action: he was supposed to hold her hand not her head
|
||||
2. Wrong details: the smoke was rising instead of drifting sideways
|
||||
3. Advisor note: motion keywords are underrepresented in the suggested prompt
|
||||
```
|
||||
|
||||
The advisor uses an explicit Refiner system prompt and receives the current prompt, the prompt that caused the rating, the original user intent, rating label, missing/wrong axes, phrase analysis, and rule-based repair candidates.
|
||||
|
||||
In `Repair prompt`, the advisor can also suggest a repaired negative prompt when the previous rating indicates wrong content, an awful result, or strong negative reward. Negative prompt advice is validated so requested subjects, actions, locations, and visual intent are not moved into the negative prompt.
|
||||
|
||||
**advisor_thinking**: Enables thinking mode for compatible Gemma CLIP text generators when the advisor is active.
|
||||
|
||||
## Appearance Safety
|
||||
|
||||
Refiner V2 treats image-to-video appearance as image-owned by default. Learned appearance, subject/character, and environment/background phrases are not auto-added by Prompt Repair or Lucky unless they are already present in the current prompt.
|
||||
|
||||
Prompt Repair is only for motion, camera work, small non-identity details, quality, and style. It will not add clothing, body traits, character identity, or background phrases from memory.
|
||||
|
||||
Legacy Void memory also skips appearance, subject/character, and environment/background tokens so old token-bank preferences cannot reintroduce those concepts.
|
||||
|
||||
## Refinement Key Loader
|
||||
|
||||
`FunPack Refinement Key Loader` lists existing V2 keys, creates a typed key when none is selected, and exposes `Import`, `Export`, and `Refresh` buttons in the browser UI.
|
||||
|
||||
Connect its `refinement_key` output to Refiner V2's `refinement_key_input` and to `FunPack Apply LoRA Weights`' `refinement_key_input` so both nodes use the same memory key.
|
||||
Entries from `Only diagnostics` runs are stored as `Advisor note:` and carry forward into `Full` mode runs. Both advisor passes see the full history so patterns across runs influence the analysis and repair.
|
||||
|
||||
## Outputs
|
||||
|
||||
**modified_positive**: Refined positive conditioning.
|
||||
|
||||
**status**: Short runtime summary.
|
||||
**status**: Short runtime summary including mode, advisor status, and repair actions taken.
|
||||
|
||||
**training_info**: Diagnostics including encoded prompt, category classification, automatic adaptation strength, Lucky memory status, and LoRA suggestions.
|
||||
**training_info**: Full diagnostics including encoded prompt, category classification, adaptation strength, Lucky memory status, and LoRA suggestions.
|
||||
|
||||
**loss_graph**: V2 learning loss graph.
|
||||
|
||||
**modified_negative**: Repaired negative conditioning. If the negative prompt is empty, no stored poor-rated tags are available, or `CLIP` is not connected, this output is empty.
|
||||
**encoded_prompts**: The exact prompt text used for encoding, plus advisor context when the advisor ran:
|
||||
|
||||
**encoded_prompts**: One string showing the exact prompt text used for positive and negative encoding:
|
||||
```
|
||||
Positive prompt: ...
|
||||
|
||||
`Positive prompt: ...`
|
||||
Advisor suggestion (applied): ...
|
||||
|
||||
`Negative prompt: ...`
|
||||
Advisor analysis: ...
|
||||
|
||||
## Vision Context
|
||||
Pre-advisor prompt: ...
|
||||
```
|
||||
|
||||
Vision inputs are intentionally advisory in this version. The refiner stores source-image size, aspect bucket, image fingerprint, changed-image status, and CLIP Vision tensor summaries in the V2 state so ratings can be interpreted with image context. It does not directly mix CLIP Vision tensors into positive conditioning.
|
||||
`Advisor suggestion` is labelled `(applied)` when validation passed and the prompt was changed, or `(rejected)` when validation blocked it. `Pre-advisor prompt` only appears when the suggestion was applied and the prompt changed. `Advisor analysis` only appears when the analysis pass ran (`Full` mode).
|
||||
|
||||
When `advisor_mode` is active and `source_image` is connected, V2 also passes that image to compatible CLIP text generators so the advisor can use visual context while still preserving the user's prompt and intent.
|
||||
## Appearance Safety
|
||||
|
||||
## Negative Repair
|
||||
Refiner V2 treats image-to-video appearance as image-owned. Learned appearance, subject/character, and environment/background phrases are not auto-added by Prompt Repair or Lucky unless already present in the current prompt.
|
||||
|
||||
When ratings mark a prior run as `Awful`, `Wrong details`, `Wrong action`, or otherwise strongly negative, V2 stores matching poorly rated prompt phrases as negative repair tags. Future runs append the strongest stored tags to `negative_prompt` before encoding `modified_negative`.
|
||||
When `feedback_prompt` is connected, appearance-related changes requested by the user are applied without restriction.
|
||||
|
||||
Positive prompt repair still uses learned phrase memory and intent alignment. Repaired phrases preserve stopwords and phrase text, so phrases like `running through the street` stay intact instead of being collapsed to only content tokens.
|
||||
## Refinement Key Loader
|
||||
|
||||
`FunPack Refinement Key Loader` lists existing V2 keys, creates a typed key when none is selected, and exposes `Import`, `Export`, and `Refresh` buttons in the browser UI.
|
||||
|
||||
Connect its `refinement_key` output to Refiner V2's `refinement_key_input` and to `FunPack Apply LoRA Weights`' `refinement_key_input` so both nodes share the same memory.
|
||||
|
||||
## Removed From V2
|
||||
|
||||
V2 does not include sigma refinement, latent refinement, manual scheduler modes, or feedback questions.
|
||||
|
||||
Scheduler behavior is automatic: good streaks make V2 gentler, bad streaks make it push harder away from bad directions.
|
||||
|
||||
## LoRA Type Rename
|
||||
|
||||
V2 uses `action` instead of `concept` for motion/action LoRA intent. Old `concept` LoRA rows are treated as `action` aliases.
|
||||
Scheduler behaviour is automatic: good streaks make V2 gentler, bad streaks make it push harder away from bad directions. In `Prompt only` mode, streaks and reward EMA are not updated so prompt-only ratings do not affect conditioning adaptation strength.
|
||||
|
||||
Reference in New Issue
Block a user