🌆DepthFM Literative

The 🌆DepthFM Literative node is the multi-image sibling of the single-shot DepthFM node, and despite the name - "Iterative" - it doesn't iterate in any special way. The typo is baked into the display name; we all just live with it. It runs the exact same flow-matching solve, but loops over every image in the batch you feed it instead of assuming one. Give it a stack of frames and you get a stack of depth maps back, one per input.

The inputs mirror the single-image node: model (a DepthFMMODEL from the 🌆DepthFM ModelLoader), image (an IMAGE batch), steps, and ensemble_size. The output is one IMAGE - the concatenated depth batch. The settings carry over too: steps 2–4 for normal use, up to 10 when you want fine detail, ensemble 2–4 as the sweet spot. One mechanical detail worth knowing: the ensemble averaging in the underlying model only supports a batch of one, and this node cheats around that by processing each frame independently, so the same defaults just work.

Why you'd reach for it: the pack's headline demo is video. Drop a clip in through a VHS or video loader that outputs an IMAGE batch, get per-frame depth out, and run it into depth-based ControlNet or parallax tricks for a consistent style pass or a 2.5D pan. It also handles the plain multi-image case - several stills measured in one queue without chaining separate DepthFM nodes.

The honest caveat: this is per-frame depth, not video depth. Each frame is estimated independently, and independent estimates can flicker between frames - the same failure the community flags for any image-model-on-video approach, and the reason DepthCrafter exists for temporal stability. DepthFM's official repo claims some temporal friendliness baked into its training, but this wrapper does zero temporal smoothing of its own, so judge a clip's stability before committing to a long render. That's a real thing people hit: a gorgeous single-frame result that shimmers on playback.

Same install and same gotchas as the rest of the pack: git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-DepthFM into custom_nodes, pip install -r requirements.txt (just torchdiffeq), restart, and the manual depthfm-v1.ckpt download into ComfyUI/models/depthfm/ - no auto-download. Frames must be a multiple of 64; the loader is CUDA-only; and VRAM now scales with your batch, so a long clip means processing in chunks unless your card is generous.

For stills, the single-image node is the cleaner pick. For a frame sequence where you need a matching set of maps in one go, this is the one you want.