FunPack Video Refiner V2

FunPack Video Refiner V2 is the heart of the pack's self-improving loop, and it's unlike anything you'll find in most custom node packs. It takes a prompt, encodes it through your CLIP, and learns from your ratings - "Missing action", "Wrong details", "Perfect" - then rewrites the prompt and steers the conditioning so the next generation gets closer to what you actually wanted. It's a feedback loop where your thumb is the training signal.

How it works

The required inputs are the two you'll touch every run: positive_prompt (the text to encode and refine) and rating (the 25-option feedback enum - the samples show the real labels: Perfect, Nailed it, Missing details, Missing action, Missing quality, Wrong details, Wrong action, Wrong appearance, -Just forget it-). The docs spell out the mapping: Missing details nudges detail phrases, Missing action boosts action/character phrases, Wrong appearance means the video got polluted by remembered character traits that weren't requested, Perfect reinforces what worked.

mode picks how much of the machinery runs: Refine (apply learned prompt shaping + conditioning adaptation - the default), Prompt only (shape the prompt, pass conditioning vectors through untouched), or Learning (record and update memory but pass everything through unchanged - the safe way to build memory before it starts steering).

advisor_mode (default Off) is the LLM advisor: Off, Only diagnostics, Only prompt, or Full. Full runs two passes - first analyze what's wrong, then repair the prompt - using the advisor_clip model (from FunPack Advisor LLM). The advisor gets four explicit inputs: original user intent, last prompt, your rating, and the feedback history. feedback_prompt is the "highest priority" override: "he was supposed to hold her hand, not her head" gets obeyed exactly, bypassing the axis logic.

The outputs that matter

• modified_positive (CONDITIONING) - the refined conditioning. This is what feeds your sampler.
• model (MODEL) - optional input/output: with an LTX-2/LTXAV model connected it does attn2 direction injection and temporal RoPE tricks. Wire it to a non-LTX model like Wan and it passes through unchanged - this feature is LTX-only.
• loss_graph (IMAGE), training_info (STRING), encoded_prompts (STRING) - diagnostics. encoded_prompts shows the exact prompt used plus advisor output, and is where you see "Advisor suggestion (applied)".
• video_latent, status, steer_mode/absolute_strength - steer_mode relative (default) finds the best conditioning for this prompt; absolute pulls toward a global learned taste; both layers them.

The details that matter

split_by_transitions (default off) is the scene-chain bridge: it detects transition words in the prompt and returns one conditioning entry per scene for FunPack LTXAV Scene Chain Sampler. Keep the character description before the first transition - it becomes the shared anchor prepended to every scene. im_feeling_lucky composes a prompt from the learned phrase memory instead of using your text. temporal_style (natural/auto/accelerate/decelerate/loop/freeze/pulse) manipulates frame-rate RoPE to change perceived motion timing - a genuinely unusual control. reset_session wipes the key's state.

Install

It's part of ComfyUI-FunPack:

cd ComfyUI/custom_nodes
git clone https://github.com/olivv-cs/ComfyUI-FunPack
pip install -r requirements.txt

or ComfyUI Manager → search "ComfyUI-FunPack". Needs transformers >= 5.0; avoid the optional hpsv3 package, which pins an older transformers and breaks the advisor.

The honest take

V2 replaced the original FunPack Video Refiner and keeps fresh state - no migration, old histories are gone. It also dropped sigma/latent refinement and scheduler controls in favor of the automatic streak logic (good streaks make it gentler, bad streaks push harder). It's the most powerful single node in this pack and the most opinionated: read encoded_prompts when a result surprises you, keep one refinement key per project, and expect the first few runs to be mostly learning before the loop visibly pays off.