Frame Reverse Module (Bending, Experimental)

There's a whole family of temporal bending nodes in this pack that share one idea - the model sees the clip as a stack of latent frames, and you can move content around inside that stack. Frame Reverse Module is the bluntest of them: flip the stack.

It has literally no inputs. No strength, no padding, no curve. It outputs a BENDING_MODULE that reverses the order of the latent frames it's given. Everything else about the pack's flexibility is gone, and that's the honest description of this node.

What it actually does

Under the hook it's one line - torch.flip on the time axis - and it only does anything when the tensor is 5-D (batch, channels, frames, height, width). Hand it a 4-D image tensor and it returns the input untouched. So this is a video-only module in the strictest sense: no spatial trick, no grid inference, nothing to configure.

What the model does with a reversed stack is the interesting part. Video DiTs like WAN are trained on clips that run forwards, with a causal 3D VAE whose temporal compression assumes temporal order. Feed one a reversed sequence and there are two plausible outcomes, both observable:

• Motion arcs invert. Something that was rising falls. Camera pushes become pulls. This is exactly what you'd get by reversing the decoded clip, so if that's all you want, do it after the VAE and save the VRAM.
• The model fights back. Because the reversed stack no longer looks like a plausible clip to the transformer, later layers can produce something that reads like a scene playing in reverse with intent - reversed motion and normal-looking content. That's the reason to use it here rather than in a video editor, and it's also the reason the results are inconsistent.

Because it's a module, it fixes nothing about when it applies. Put it as-is on double:10-20 and every step sees reversed frames, so the whole denoise happens in reverse-time-land. Wrap it in Timestep Gated Bending and you can reverse during the composition steps and then let the sampler work in normal time for the detail steps - which is where the "reversed motion, coherent look" version tends to come from.

Output and wiring

Output: BENDING_MODULE, and that's the entire interface - no inputs at all.

Wire it into DiT Block Bending (with the video model's block spec, spatial on) or Attention Map Bending (it reverses the order of the frames of the attention map, so what each prompt token attends to at each frame comes from the opposite end of the clip). Both are meaningful; the attention route is the one the pack's temporal workflow explores.

Install

ComfyUI Manager → search "ComfyUI Model Bending", or:

cd ComfyUI/custom_nodes
git clone https://github.com/abuzreq/ComfyUI-Model-Bending
pip install -r ComfyUI-Model-Bending/requirements.txt

kornia and scikit-learn, then restart ComfyUI. Model weights are on you - the pack ships none, and the video workflows link their own WAN downloads.

Where people get burned

• Using it on images. Silent no-op. Nothing warns you, because from the module's point of view there was nothing to reverse.
• Expecting a tunable effect. This is the one module in the pack with zero knobs. If you want some reversal over time, that's Frame Ramp with a negative w_end on a different module, not this node.
• Forgetting the experimental flag. The author validated these temporal modules on tiny random-weight WAN models in the test suite, not on real weights at real resolutions. Treat the first few tries as reconnaissance.
• Assuming frame count changes. It reverses order, not length. The clip is the same length and the same fps; only the direction of what the model believed changes.
• Stacking two of them. Reversing twice is a no-op that costs you an extra blend pass and a lot of confusion.