Outverse SD3 Model Pred

SD3 and SD3.5 are flow-matching models, but they don't drop into FlowEdit editing the way Flux does. FlowEdit wants a model that, handed an image latent, predicts its denoised output directly - not one that expects to start from pure noise. OutSD3ModelSamplingPred is the adapter: a MODEL-in, MODEL-out patch you slip between your checkpoint loader and the rest of the FlowEdit pipeline. It rewrites how the model samples so this pack's inversion-free edit sampler can run on it at all. Nothing visible changes on its own - skip it and the FlowEdit nodes downstream produce garbage.

The mechanism, straight from the code: it clones your model and swaps in a custom ModelSamplingAdvanced built on ComfyUI's ModelSamplingDiscreteFlow (the flow-matching sigma schedule) fused with one of two prediction classes. The default, OutCONST, does the standard input - output·sigma denoise cleanup - but its noise_scaling returns your latent untouched instead of mixing in noise. That's the whole "outverse" idea: stock SD3 sampling assumes a noise-to-image trip, while FlowEdit builds its own interpolated latents ((1-sigma)·image + sigma·noise) inside the sampler. The model needs to treat whatever latent it's handed as an image, and this patch makes it.

The inputs are short, so here's all of them:

• model - any SD3/SD3.5 checkpoint, official or fine-tuned.
• model_shift - default 3. The flow-matching knob that spreads sampling effort between composition and detail. 3 is SD3.5's native value and the shipped workflow keeps it; nudge toward 5+ if your edit comes out mushy on details, toward 1–2 if the composition wobbles.
• reverse_ode - default off. When on, the patch uses ReverseCONST, where the model's raw prediction is used as the denoised result - the reversed direction the source→target FlowEdit map solves in. The example workflow ships with it on; leave it there unless you know why you're flipping it.
• Output: the patched MODEL, which feeds both BasicScheduler (for sigmas) and the pack's FlowEditCFGGuider.

Install

It's one of three nodes in the ComfyUI_SD3_Flowedit pack, so install once and all three appear. ComfyUI Manager: search "ComfyUI_SD3_Flowedit" and install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/raindrop313/ComfyUI_SD3_Flowedit

then restart ComfyUI. No model downloads, no heavy Python deps - requirements.txt lists only a tiny einsum helper the code never imports. The model is whatever SD3/SD3.5 checkpoint you already have in models/checkpoints; the author's example loads SD3.5 Medium and notes you can fine-tune for your own edit distributions. The pack is a port of logtd's ComfyUI-Fluxtapoz, so the algorithm behind it is well-trodden.

Common issues

• Only use it on SD3/SD3.5. It's built on the flow-matching schedule, so patching an SDXL or SD1.5 model with it produces nonsense. If you're here to edit with a different model family, you want the Flux version of this pack (Fluxtapoz), not this one.
• Feed both consumers the patched model. In the example, the patched output drives BasicScheduler and the guider. Wire the scheduler's model input to the raw loader output instead, and your sigmas won't match the model's schedule - the classic "everything comes out noisy" failure.
• Check reverse_ode first. If your result looks like a noised smear rather than an edit, the direction flag is the prime suspect. The shipped workflow runs it on.
• It's a patch, not a predictor. The node name promises prediction, but nothing generates here - the real work happens in FlowEditSampler and FlowEditCFGGuider, both in the same pack. Install all three, or this does nothing visible at all.