ComfyUI Node
Difforum · Restyle (Deforum / AnimateDiff look)
Give a video-model render the Deforum / AnimateDiff / Disco look. MiniMax H3 or LTX supplies the motion; an image model (SDXL, SD1.5, Flux, turbo distills) re-paints every frame in your look, with the previous stylized frame carried along the video's optical flow and mixed in. That feedback is what made Deforum morph and AnimateDiff boil - here it rides on the video model's motion instead of a synthetic camera. Styles: clean restyle (steady), animatediff boil (texture re-rolls each frame), deforum morph (strong feedback smear), disco flicker (high denoise, per-frame seed, loose colour). `custom` uses feedback / seed_mode / color_hold as set. Connect a Director to use its scene prompts and energy.
Difforum · Restyle (Deforum / AnimateDiff look)
- video
- model
- positive
- negative
- vae
- direction
- prompts
- frames
- report
◄styledeforum morph►
◄denoise0.45►
◄steps12►
◄cfg5.0►
◄sampler_nameeuler►
◄schedulernormal►
◄cadence1►
◄long_edge1024►
◄seed0►
◄feedback0.35►
◄seed_modefixed►
◄color_hold0.50►
◄follow_energytrue►
◄flow_scale0.50►
CategoryDifforum/4 · Render
Inputs (21)
| Name | Type | Default | Description |
|---|---|---|---|
| video | IMAGE | Frames from H3 / LTX / any video (Get Video Components). | |
| model | MODEL | An image model in the look you want. Turbo SDXL / DMD2 at 4-6 steps is fast and good. | |
| positive | CONDITIONING | The look, e.g. 'oil painting, thick brush strokes'. | |
| negative | CONDITIONING | Negative conditioning (CLIP Text Encode). | |
| vae | VAE | The VAE that matches the model. | |
| style | COMBO | deforum morph | deforum morph: strong feedback, smears along motion. animatediff boil: texture re-rolls each frame. disco flicker: high energy, loose colour. clean restyle: steady painted version. custom: uses feedback / seed_mode / color_hold below. |
| denoise | FLOAT | 0.450.05–1 | How much each frame is re-painted. 0.3 keeps the render, 0.6+ re-imagines it. |
| steps | INT | 121–100 | Full-denoise steps; each frame runs steps x denoise. |
| cfg | FLOAT | 5.00–30 | Prompt strength. Turbo models: 1-2. |
| sampler_name | COMBO | euler | Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo. |
| scheduler | COMBO | normal | Noise schedule. karras / normal for regular models, sgm_uniform for turbo. |
| cadence | INT | 11–8 | Re-paint every Nth frame; the rest follow the flow. 2 = twice as fast. |
| long_edge | INT | 10240–4096 | Working size. 0 = the video's size. Upscale afterwards for 2K. |
| seed | INT | 00–4294967295 | Noise seed. |
| directionopt | DIFFORUM_DIRECTION | Director: scene prompts (with CLIP) and the energy curve. | |
| promptsopt | DIFFORUM_PROMPT | Prompt travel; otherwise the Director's scene prompts, otherwise positive. | |
| feedbackopt | FLOAT | 0.350–0.95 | custom: share of the previous stylized frame in the next one. |
| seed_modeopt | COMBO | fixed | custom: per frame = texture re-rolls every frame (boil / flicker). |
| color_holdopt | FLOAT | 0.500–1 | custom: how much each frame keeps the source colours. |
| follow_energyopt | BOOLEAN | true | With a Director: denoise follows its energy curve (0.5 = as set). |
| flow_scaleopt | FLOAT | 0.500.25–1 | Resolution of the motion estimate that carries the previous frame. 0.5 = fast. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| frames | IMAGE | — |
| report | STRING | — |