Difforum

Direct the camera. Let any model render it.

A timeline for ComfyUI: draw scenes, camera moves and energy with the mouse, then render the shot with MiniMax H3, LTX-2, or the classic Deforum feedback look. The camera also goes out to After Effects and Blender.

Video models render well, but you steer them with words. Difforum gives you back control over the camera and the timing:

1. Direct. In the Director node you lay out scenes (prompt + mood), camera moves (picked from 25 visual presets), keys (exact frames where something must happen) and an energy curve. Press ▶ to preview the move; the preview uses the same engine that renders. The mouse wheel zooms the timeline, so a 60-second shot is as easy to edit as a 5-second one.

2. Previz. Every template has an Animatic: the whole shot, with timecode, prompt, camera move, energy and keys burnt in, in a few seconds. The Director's Previz only button mutes the render outputs, so Queue runs just the previz. Click it again to render.

3. Render. One direction wire goes to the renderer of your choice:

   Renderer	Difforum node	You get
   MiniMax H3	H3 Guides / H3 Shot	Your camera turned into keyframes, first/last frames and a camera prompt, with no hand-placed frames. Areas the camera uncovers are painted by Fill Reveal (AI)
   LTX-2 / 2.5	LTX Guides	Keyframes inside the LTX latent
   Any image model (SDXL, Flux, SD1.5, turbo)	Feedback Sampler	The Deforum look: every frame re-imagines the last
   Realtime (turbo models, webcam)	Live Sampler	Plays inside the node, with Spout / OBS output

4. Finish. Camera Export writes the same camera for After Effects (.jsx) and Blender (.py) so titles, 3D and comp lock to the AI shot.

Pick a look once. The Director's look (cinematic, documentary, deforum_morph, animatediff_dream, disco_diffusion, vqgan_clip, flicker_experimental, psychedelic, music_video, stop_motion, hand_drawn) sets the Feedback Sampler's colour, detail and energy, and adds the matching style sentence to the H3 / LTX prompt, so the same aesthetic carries across renderers.

The Deforum / AnimateDiff look on MiniMax H3. Restyle re-paints every frame of an H3 / LTX render with an image model in your look, carrying the previous painted frame along the clip's own motion: that feedback is what made Deforum morph and AnimateDiff boil, now riding on H3's motion (deforum morph, animatediff boil, disco flicker, clean restyle). Look Mix adds grain or flicker cuts on top. Template 13 restyles any clip you already rendered, with a depth ControlNet holding the structure so denoise can go up to 0.6-0.8. Template 14 is the real AnimateDiff motion module on any video (vid2vid); template 15 is its fast AnimateLCM version with a hi-res pass.

Multikeyframing. Mark moments on the Keys track and feed one picture per key to Keyframe Images. The Feedback Sampler travels through them, H3 / LTX Guides anchor them, the Animatic flashes them. Built for installations and pieces that must hit an image on a beat.

Orchestration from outside. Shot Script writes the Director timeline from text (0s | calm | dolly_in slow | prompt), a CSV, a file in input/ or any STRING node such as an LLM. Keyframe Assets loads a folder of stills named by time (0s_wide.png, 4.5s_sky.png) as keyframes, so styled pictures made anywhere drive H3 without rendering a look pass (template 16).

No camera expressions to write. They are still there if you want them.

1. Install from ComfyUI Manager (search "Difforum"), or:

   cd ComfyUI/custom_nodes && git clone https://github.com/chillithebillis/Difforum.git difforum

2. Restart ComfyUI and open Templates → Difforum.
3. Start with 01_storyboard_no_model. It needs no model: load a picture, shape the timeline, press ▶, then Queue.

| 13 | restyle_any_video | Restyle a clip you already have (H3, LTX, live action) into the Deforum / AnimateDiff / Disco look, then 2K | a turbo model + a video | | 14 | animatediff_on_video | AnimateDiff vid2vid: SD 1.5 + AnimateDiff-Evolved + depth / canny ControlNets on any clip, then 2K | AnimateDiff-Evolved, Advanced-ControlNet, an SD 1.5 checkpoint, a motion module, SD 1.5 ControlNets | | 15 | animatediff_lcm_fast | AnimateDiff LCM: template 14 in 8 steps with AnimateLCM, plus a x1.5 hi-res pass | as 14, with the AnimateLCM motion module + LoRA | | 16 | h3_keyframe_assets | Stills + shot script → H3: a folder of styled keyframes and a text script drive H3 guides; no look pass | H3 ref2va + your stills |

Every template is laid out in the same blocks: Control (switches and notes), Direction, Previz on the top row; Models → Render → Restyle → Upscale 2K → Output on the bottom row. The Workflow Switches panel turns each block on or off (a render block is muted, a pass-through block such as Upscale is bypassed) and ⌖ jumps to it. H3 templates have a Live Preview (TAEH3) block that shows the video forming at every step with KJNodes' Model Preview Override and taeh3.safetensors in models/vae_approx. H3 templates render in two stages: H3 renders small (640 px), the H3 Latent Upscale block lifts the latent 2x with the Minimax H3 Latent Upscaler and H3 re-samples a few steps at full size, so the detail is generated by H3 and holds over time. The Upscale 2K block then finishes in pixels (RealESRGAN_x4plus or any model in models/upscale_models).

Load Image ─► Guide Frames ─► Keyframes ─► Fill Reveal (AI) ─► H3 Guides ─► Sampler ─► Video + audio
                  ▲                            ▲
Setup ─► Director (camera, scenes) ─► Camera → Prompt ─► MiniMax H3 Reference to Video

The official Multiframe Reference template anchors images you place by hand. Template 10 anchors frames that come from your camera move, so H3 performs the dolly, orbit or crane you drew. Template 08 does the same with H3's first/last-frame model, and splits clips longer than 20 s into segments. Both use ComfyUI's core MiniMax H3 nodes.

The full reference is in docs/NODES.md. Every node also shows its description inside ComfyUI.

The Feedback Sampler runs on any image model and fixes what made the original Deforum hard to use:

• Depth follows the image, so parallax stays true for the whole clip.
• 3D moves never freeze. Without a depth map they become pseudo-3D, and the node says so.
• Cadence without pops. Only every Nth frame is diffused; the frames in between are crossfaded.
• Revealed edges are repainted instead of smeared.
• Colour locks per scene, so prompt travel can change the palette.
• Steps scale with the energy, like Deforum: a frame at denoise 0.5 runs half the steps. With cadence 2 that is about 4x faster than sampling every frame in full.

pytest runs without a GPU or ComfyUI. Templates and the node reference are generated from the code (tools/). See CONTRIBUTING.md.

MIT licensed.