Nodes/ComfyUI-FunPack/FunPack Video Refiner V2
ComfyUI Node

FunPack Video Refiner V2

Prompt-owned Video Refiner V2. Encodes through the connected CLIP, learns from ratings, and writes LoRA suggestions without sigma/latent/feedback systems.

By digital-garbage·Created about a year ago·Updated 24 days ago· 16
FunPack Video Refiner V2
  • clip
  • advisor_clip
  • positive_conditioning
  • lora_stack
  • clip_vision_output
  • source_image
  • model
  • latent
  • modified_positive
  • status
  • training_info
  • loss_graph
  • encoded_prompts
  • model
  • video_latent
positive_prompt
ratingMissing action
modeRefine
advisor_modeOff
reset_sessionfalse
refinement_key_input
user_intent_prompt
im_feeling_luckyfalse
advisor_thinkingtrue
feedback_prompt
temporal_stylenatural
split_by_transitionsfalse
steer_moderelative
absolute_strength0.60
CategoryFunPack/Refinement

Inputs (22)

NameTypeDefaultDescription
positive_promptSTRING
ratingCOMBOMissing action25 options: -Just forget it-, Perfect, Nailed it, Missing details, Missing action, Missing quality, +19
modeCOMBORefineRefine applies learned prompt and conditioning changes. Prompt only shapes the prompt but passes conditioning vectors through unchanged. Learning records observations and ratings while passing everything through unchanged.
advisor_modeCOMBOOffCLIP text-generation advisor. Full: repair prompt + show diagnostic. Only prompt: repair prompt silently. Only diagnostics: report advice without changing the prompt. Off: disabled.
clipoptCLIPOptional text encoder. When connected, V2 encodes the prompt itself.
advisor_clipoptCLIPOptional separate CLIP/Gemma text generator for Advisor. If disconnected, Advisor falls back to the main CLIP.
positive_conditioningoptCONDITIONINGOptional pre-encoded Gemma3/LTX2 conditioning. Used only when CLIP is not connected.
reset_sessionoptBOOLEANfalse
lora_stackoptFUNPACK_LORA_STACKOptional stack from FunPack LoRA Loader. V2 writes prompt-specific suggested weights.
clip_vision_outputoptCLIP_VISION_OUTPUTOptional CLIP Vision output describing the source image. Stored as advisory context; it is not blended into positive conditioning.
source_imageoptIMAGEOptional original/source image or frame batch. V2 stores size, aspect ratio, and a simple fingerprint to notice changed inputs.
modeloptMODELOptional model. LTX-2 / LTXAV family only - attn2 direction injection, attention temperature, temporal RoPE, and anchor transfer all require LTX transformer internals. Connecting a non-LTX model (e.g. WAN) passes it through unchanged. Connect the model output to your sampler.
refinement_key_inputoptSTRINGOptional linked refinement key, for example from FunPack Refinement Key Loader. Overrides the refinement_key widget when connected.
user_intent_promptoptSTRINGOptional raw user/request prompt. V2 can use this as an intent source to repair missing action/detail/quality phrases if an enhancer omits them.
im_feeling_luckyoptBOOLEANfalseCompose a learned prompt from V2 phrase memory, then encode it through the connected CLIP.
advisor_thinkingoptBOOLEANtrueLet compatible CLIP text generators use thinking mode for advisor diagnostics and prompt repair.
feedback_promptoptSTRINGOptional user feedback describing what was specifically wrong with the previous output (e.g. 'he was supposed to hold her hand, not her head'). Has highest priority in the advisor system prompt.
temporal_styleoptCOMBOnaturalControls how the model perceives motion timing via frame_rate RoPE manipulation. natural=no change, auto=per-scene director picks motion energy from each scene's prompt (needs the Scene Chain Sampler), accelerate=faster motion, decelerate=heavier motion, loop=circular temporal coords, freeze=highly compressed time, pulse=repeated ease-down motion holds per scene (needs the Scene Chain Sampler).
split_by_transitionsoptBOOLEANfalseDetect transition words and return one conditioning entry per scene for FunPack LTXAV Scene Chain Sampler. Leave off for normal single-conditioning workflows.
latentoptLATENTOptional video latent for creativity masking. Takes priority over any saved latent for this key. Connect your i2v or previous KSampler output here.
steer_modeoptCOMBOrelativeRelative: per-prompt steering (finds the best conditioning for THIS prompt — default). Absolute: pull conditioning toward the global learned taste, regardless of prompt. Both: layer them.
absolute_strengthoptFLOAT0.600–2How hard Absolute mode pulls toward the global taste direction. 0.6 is visible but non-destructive; raise for stronger override of the prompt.

Outputs (7)

NameTypeDescription
modified_positiveCONDITIONING
statusSTRING
training_infoSTRING
loss_graphIMAGE
encoded_promptsSTRING
modelMODEL
video_latentLATENT