Flow Matching Stage

The Flow Matching Progressive Upscaler does everything in one node: upscale, re-noise, denoise, skip-blend, repeat. FlowMatchingStage is the same machine, but as a single stage you chain by hand - one node per doubling - and the reason that's worth doing is ComfyUI's caching. When you change the settings of a later stage, ComfyUI can short-circuit the unchanged earlier stages and skip straight to what you edited. Tuning your final upscale pass no longer means re-running the whole ladder.

It does exactly what one stage of the progressive upscaler does: takes a latent, applies scale_factor, injects flow-consistent re-noise (noise_ratio), runs the sampler, and blends the skip residual back in (skip_blend). The mechanism is the same "rewind the clock" logic - re-noise puts the latent back onto the flow trajectory so the model refines detail instead of inventing a new composition - just exposed per stage instead of baked into a loop.

Inputs that matter:

• model, positive, negative, latent - standard sampler wiring.
• seed - controls the re-noising for this stage (and typically the sampler noise).
• steps (16), cfg (4.5), sampler_name, scheduler - the sampler for this stage. Flow models want the Euler family; this isn't the place for Karras.
• scale_factor (1.0) - how much this stage resizes. 1.0 means no scaling, which is a legitimate choice for a pure refinement stage.
• noise_ratio (0.0) - how much flow-noise to inject (0 = keep latent, 1 = pure noise). This is the knob that corresponds to "how far back on the trajectory."
• skip_blend (0.5) - 0 = all denoised, 1 = all pre-sampler latent. High early, low late.
• denoise (1.0) - strength supplied to the sampler.
• upscale_method (bicubic) - resize kernel. Note lanczos is a trap: it falls back to bicubic because the PIL-based path damages latents.

Optional controls: enable_dilated_sampling (default disable here, unlike the progressive node) adds the dilated refinement lap, with dilated_downscale, dilated_blend (frequency-domain blend), dilated_min_steps (raise to 4+ for lightning models), dilated_seed_mode, and dilated_denoise. And reduce_memory_use (default enable) avoids extra tensor clones to save VRAM - worth knowing when you're stacking stages at 4K.

Outputs. latent is the refined result; presampler_latent is the latent before sampling, which you can route into ComfyUI's Custom Sampler nodes if you want to take over the sampler yourself; next_seed chains to the next stage's seed for deterministic runs; and model, positive, negative pass through.

Install. Same pack - ttulttul/ComfyUI-FlowMatching-Upscaler. ComfyUI Manager, search "ComfyUI-FlowMatching-Upscaler", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/ComfyUI-FlowMatching-Upscaler

No model downloads; deps are numpy, torch, einops, aiohttp - all already in ComfyUI. It registers under latent/upscaling.

Gotchas. The naming overlaps with the Progressive Upscaler on purpose, and the default enable_dilated_sampling is flipped to disable here - if you expect dilated refinement on by default, you'll silently skip it. Also remember this node keeps the LOW_VRAM fallback that the Prep/Merge pair deliberately drops, so if you're on tight memory, prefer chaining these over the modular custom-sampler path.