Flow Matching Progressive Upscaler

This is the headline node of the ttulttul/ComfyUI-FlowMatching-Upscaler pack, and it's built around a genuinely different idea: instead of stretching your latent and blasting random noise over it (the standard img2img upscale move), it rewinds the clock. You finish a low-res image at flow time t=0; the node resizes the latent up, then re-noises it back to somewhere like t=0.6 - back onto the model's flow trajectory - and lets the sampler walk forward again at the higher resolution. Same journey, bigger canvas. Because it resumes an existing path rather than starting fresh, it fills in high-frequency detail instead of hallucinating a new composition.

The author (ttulttul, who also did the damofusion-style ComfyUI-Iterative-Mixer) ships this in a pack that now focuses on exactly this: the progressive upscaler, its modular stage nodes, the DyPE patchers, and a latent upscaler. Everything flow-matching, nothing else.

Why you'd reach for it. Flow models like Qwen Image and Flux2 don't obey the old SDXL rules - they expect a straight-line trajectory, and old advice like "denoise 0.35 with Karras" actively fails on them (Karras concentrates sampling in the middle of the path, which distorts a straight flow). This node handles the re-noising the flow-consistent way and gives you the knobs in one place: it's the "generate at native res, then climb" strategy from the upscaling playbook, but staged so each doubling gets its own denoise pass. And upscaling is genuinely optional here - set total_scale to 1.0 and it becomes a pure refinement pass, a second chance to add detail without changing resolution.

How it works, stage by stage. Each stage: (1) upscale the latent, (2) re-noise with a linear mix of fresh noise (renoise_start 0.35 tapering to renoise_end 0.15), (3) denoise with your sampler, (4) blend the skip residual back in to hold composition (skip_blend_start 0.8 → skip_blend_end 0.05), and (5) optionally run a dilated refinement lap - downscale, short sample pass, and blend back in the frequency domain (FFT-based, so it pulls low frequencies from the dilated pass and keeps high-frequency detail from the base). That dilated pass is your weapon against the over-sharpened "hallucinated pores" artifact. There's also a cleanup_stage toggle for a final no-scaling polish pass. If you run out of VRAM mid-stage, it automatically drops into a LOW_VRAM streaming fallback - slower, but it won't crash. Inpaint masks ride along and get resized in lockstep, so masked workflows stay aligned.

Inputs that matter:

• model, positive, negative, latent - the four you already know from any sampler.
• seed - base seed controlling per-stage re-noising.
• steps_per_stage (16), cfg (4.5), sampler_name, scheduler - per-stage sampler settings. For flow models, stick with the Euler family; avoid the classic DDPM-era scheduler choices unless you're testing.
• total_scale (4.0) and stages (2) - where you're going and how many doublings to get there.
• upscale_method (bicubic) - the inter-stage resize kernel.

The noise_schedule_override and skip_schedule_override strings let you hand-write per-stage values as comma-separated lists; denoise is a global multiplier. Skip those until you're chasing a specific artifact.

Outputs. latent is the result, next_seed chains deterministically into another stage, and model, positive, negative pass through so you can keep wiring downstream.

Install. ComfyUI Manager, search "ComfyUI-FlowMatching-Upscaler", install, restart. Or:

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

Dependencies are numpy, torch, einops, aiohttp - nothing heavy, no model downloads. Restart and it's under latent/upscaling.

Gotchas. The pack's own README warns that noise isn't random here - it's what makes the latent statistically valid at that timestep, so fight the urge to crank renoise_start down to zero (you get waxy, plastic output). And if you're on a lightning/distilled flow model that samples in 4 steps, raise dilated_min_steps - the default dilated pass may be too short to do anything. Start with defaults, get one clean upscale, then tune.