Difforum · Fill Reveal (AI)

Parallax has a hard limit that every "living photo" demo quietly papers over: the depth map only describes the original camera view. Move the virtual camera and you reveal background that was never photographed. There is no pixel data there - a depth warp can stretch the border, or leave gray, and neither is scenery.

Fill Reveal is the honest fix. Guide Frames, H3 Shot and Keyframes mark the invented area in a mask; this node repaints only that area with an image model, and keeps the known pixels exactly.

How it does it

Per frame: take the mask (grow it a little into the known image so the seam hides, feather the edge), hand the frame plus mask to ComfyUI's core InpaintModelConditioning, run a normal KSampler on it, decode, then composite the result back through the feather with a soft alpha. seed + frame index seeds each frame differently, so a multi-frame fill doesn't paste the same invention into every hole.

It runs on plain MODEL/VAE/CONDITIONING, which means any image model works. The cleanest seams come from a dedicated inpaint checkpoint - SDXL inpainting, Flux Fill - because those models are trained to read the mask as a condition rather than just "generate here", and the difference shows up exactly at the edges. A regular checkpoint works too; it just leaves more seam.

The inputs that matter

• images and masks - the frames with holes, and the mask where 1 means "fill this". That's Guide Frames' default convention, so they wire straight across.
• frames - all, first, last, or first + last. This is the practical dial. Run first + last for an FL2VA shot, all for a keyframe set that's about to guide a model. It's one diffusion per frame, so the tip in the tooltip is the whole cost model: only paint the frames a video model will actually see.
• positive / negative - describe the scene so the fill matches. Inpainting follows the prompt; if the fill is inventing a shopping mall behind your forest, that's a prompt problem, not a node problem.
• grow (16) and feather (12) - the seam controls. Grow pushes the mask into the known image so the model redraws a little of the real photo as it blends; feather softens where painted meets original. These are the two numbers to fiddle with, and the classic values (4–12 px of feather for a still image) hold up here.
• steps, cfg, sampler_name, scheduler - same rules as everywhere else in the pack. Turbo/LCM/DMD2 models want cfg 1–2 and an lcm or euler_ancestral sampler with sgm_uniform; a regular SDXL fill is happy at cfg 4–7.

Outputs

images (the frames, with holes filled) and info - which is a per-frame report of how much of the image it actually filled, plus a "nothing revealed, kept" line on frames where the camera didn't uncover anything. That line matters: it means the mask was empty and the node correctly did nothing instead of doing something.

Install

ComfyUI Manager → search "Difforum", or:

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

Restart. You need an image model - the pack ships no weights - and the node uses ComfyUI's own InpaintModelConditioning, so nothing extra installs.

Where people get burned

The cost. This is real sampling, one pass per frame, at full resolution. Painting a 240-frame clip with SDXL is a long evening. Check how much is actually revealed first - info on Guide Frames reports the average revealed area - and if it's 8%, consider letting the feedback loop or the video model handle it and save Fill Reveal for the frames that get anchored.

The other one is edge quality. A big grow with a small feather gives you a hard ring where the model's invention meets your photo; a small grow with a big feather bleeds the fill into the parts you wanted preserved. Start at the defaults, look at one frame at 100%, and adjust from what you see - the composite is deterministic, so a single test frame tells you everything.