Nodes/Difforum/Difforum · Feedback Sampler
ComfyUI Node

Difforum · Feedback Sampler

The Deforum look on any image model: every frame is the last one, moved by the camera and re-imagined by the sampler. Plug a Director `direction` wire and a first frame; that is the whole setup. Works with SD1.5, SDXL, Flux, SD3.5 and turbo/LCM/DMD2 distills (plain MODEL / VAE / CONDITIONING). Connect a depth map (e.g. Depth Anything V2 on the first frame) and set the Director to 3d for real parallax - the depth then follows the image frame by frame. Outputs the frames, the tracked depth per frame (for comp / relight / video-model guides) and a run report.

By chillithebillis·Created 3 months ago·Updated about 7 hours ago· 8
Difforum · Feedback Sampler
  • model
  • positive
  • negative
  • vae
  • init_image
  • direction
  • options
  • params
  • camera
  • strength
  • cfg_curve
  • prompts
  • depth
  • control_net
  • control_image
  • key_images
  • frames
  • depth
  • report
◄steps20►
◄cfg6.0►
◄sampler_nameeuler►
◄schedulernormal►
◄cadence1►
◄control_strength0.60►
◄energy0.50►
◄key_indices—►
◄key_pull0.65►
◄key_approach12►
CategoryDifforum/4 · Render

Inputs (26)

NameTypeDefaultDescription
modelMODELThe image model (SD1.5 / SDXL / Flux / turbo). MODEL output of your checkpoint or LoRA chain.
positiveCONDITIONINGLook / style prompt. Used for every frame when no scene prompts reach the sampler.
negativeCONDITIONINGNegative conditioning (CLIP Text Encode).
vaeVAEThe VAE that matches the model.
init_imageIMAGEThe first frame. Everything grows from this image.
stepsINT201–200Steps for a full denoise. With step scaling (default) a frame at energy 0.5 runs half of them, like Deforum.
cfgFLOAT6.00–30Prompt strength. Turbo / LCM / DMD2 models want 1-2; regular SDXL 4-7.
sampler_nameCOMBOeulerSampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
schedulerCOMBOnormalNoise schedule. karras / normal for regular models, sgm_uniform for turbo.
cadenceINT11–12Diffuse every Nth frame; the frames in between ride the camera and crossfade. 2 is almost invisible and twice as fast.
directionoptDIFFORUM_DIRECTIONThe Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
optionsoptDIFFORUM_OPTIONSDifforum · Render Options, for manual control.
paramsoptDIFFORUM_PARAMSFrom Setup: frame count, fps, width and height. Plug the same Setup everywhere.
cameraoptDIFFORUM_CAMERAA camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
strengthoptDIFFORUM_SCHEDULEA schedule that replaces the Director's energy curve (denoise per frame).
cfg_curveoptDIFFORUM_SCHEDULEA schedule for cfg per frame.
promptsoptDIFFORUM_PROMPTPrompt travel from Prompt Travel (or the Director with CLIP connected).
depthoptIMAGEDepth of the first frame (white = near), or a per-frame batch.
control_netoptCONTROL_NETOptional ControlNet (depth, canny...) applied on every diffused frame.
control_imageoptIMAGEPer-frame control images; empty = the frame itself.
control_strengthoptFLOAT0.600–3ControlNet strength.
energyoptFLOAT0.500–1Denoise per frame when no Director or strength curve is connected. It is how much each frame is re-imagined: 0.3 = steady, 0.5 = classic Deforum morph, 0.7+ = the image changes every few frames.
key_imagesoptIMAGEKeyframe Images: the travel steers toward each picture and lands on it at its frame. This is multikeyframing.
key_indicesoptSTRINGFrame numbers of the key pictures (Keyframe Images 'indices').
key_pulloptFLOAT0.650–1How strongly the frame is pulled toward the next key picture before diffusion. 1 = arrives exactly on it, 0.6 = strong resemblance, 0.3 = a hint.
key_approachoptINT121–240Frames before a key over which the pull ramps up. Short (6-12) = sudden arrival, long (24-48) = the image slowly becomes the key. Long shots want long approaches.

Outputs (3)

NameTypeDescription
framesIMAGE—
depthIMAGE—
reportSTRING—