SPEED in ComfyUI: LC Speed Boost 🚀, tested on Krea 2, Z-Image, Flux.2 Klein and SDXL

[lonecatone23]

Introduction:

I'm Lonecat. I build the LC node packs for ComfyUI and the Krea 2 workflows a lot of CivitAI users run. I built SPEED into a ComfyUI node, LC Speed Boost 🚀, and tested it across the models people actually use. Here is what held up, what I changed and what I found along the way.

• Paper: Spectral Progressive Diffusion for Efficient Image and Video Generation (Xiao, Chao, Yariv, Wetzstein, 2026)
• Node: ComfyUI_LC123_nodes (LC Speed Boost and LC Speed Boost KSampler, v1.49.7+)
• Also leans on the MIT licensed ComfyUI-SPEED-SwarmNeo. Credit where it is due.

Note: every comparison is the same seed, prompt and step count, on and off. Times include VAE decode. Your numbers will differ. The ratios are what matter.

---

How it is built

• Two nodes: a sampler wrapper (any KSampler / SamplerCustomAdvanced setup) and a KSampler with it built in (start / end at step and leftover noise, so it chains like KSampler Advanced).
• Flow models (Krea 2, Z-Image, Flux.2 Klein, Qwen-Image, Wan): the spectral shrink / pad from the paper, then the state is scaled by k = r / (1 + (r - 1) t) and sampling carries on at k * t.
• SDXL / Pony / Illustrious (eps and v prediction): ComfyUI's sampler state is picture + sigma * noise, so after the same shrink / pad the state is scaled by r and sampling carries on at r * sigma.
• The switch point is set on one 0..1 scale for both: t for flow, sigma / (1 + sigma) for SDXL. It follows the schedule, not the step number.
• It steps aside where the layout comes from an input image: denoise under 0.9 and inpaint masks. The console says why.

---

Results

Speed (RTX 5090, flow models):

Model	Steps	Off	Auto	Faster
Krealism V3.1 (Krea 2)	10	10.1 s	5.4 s	1.87x
Z-Image Turbo	8	3.7 s	2.0 s	1.85x
Flux.2 Klein 9B	4	-	-	1.3x
Krea 2 two-pass (both samplers)	10	16.6 s	8.3 s	2.0x

Speed (RTX 5060 Laptop 8 GB, SDXL base, 1024², 3 seeds):

Setup	Off	Auto	Faster
dpmpp_2m karras, 25 steps	13.65 s	10.01 s	1.36x
dpmpp_2m karras, 30 steps	16.18 s	11.30 s	1.43x
euler normal, 25 steps	13.65 s	8.76 s	1.56x
832x1216, dpmpp_2m karras, 25 steps	13.26 s	9.24 s	1.43x

• VRAM: peak unchanged. The last steps run at full size and the weights take most of the card.
• Detail: skin, pores, lashes and fabric hold up. Side by side I could not pick the boosted image by detail alone.
• vs half size + latent upscale + refine: Speed Boost was faster (10 steps vs 15) and cleaner (no waxy skin or blocky patches).

---

The switch point

Flow models: auto switches at the step nearest t = 0.7.

• 0.7 kept every bit of detail on Krea 2, Z-Image and Klein.
• 0.85 and 0.9 did not bring back the original framing (the layout is settled in the first couple of steps), so switching sooner only cost time.
• 🤔 The paper's own math lands around 0.78 to 0.85 on these schedules. That is safe, just slower.
• "Nearest", not "first under": on a 9-step Krea 2 Turbo schedule one step sits at 0.706. First-under skipped it and left only 3 of 9 steps at full size. Nearest gives 4, for about a third of a second.

SDXL: auto switches at the step nearest sigma / (1 + sigma) = 0.6.

Switch at	karras 25 step	Result
0.72	9	clean
0.62	11	clean (auto)
0.56	12	clean
0.44	14	harsh freckles on Illustrious
0.31	16	broken: crosshatch noise on anime

• Tested on Juggernaut XL, SDXL base, perfectionRealistic ILXL, WAI Illustrious and CyberRealistic Pony, 2 seeds each.
• euler, euler_ancestral, dpmpp_2m_sde and dpmpp_sde, plus karras, normal, simple and sgm_uniform: all clean at auto.

---

Things I found

• Framing: the one real trade. The layout is decided at the lower resolution, so a seed frames tighter and closer. On big renders that works in its favor (no drift wide). Few-step and distilled models feel it most (a DMD SDXL checkpoint at 8 steps framed much tighter).
• Hands: a user reported fused fingers on Krea 2 Turbo. 12 seeds of hard hand poses (guitar chords, camera grip, hoodie laces, peace sign) came out clean in every mode. The culprit turned out to be a LoRA. The "nearest 0.7" rule came out of that test.
• Detail Daemon: keep it off the boosted sampler. On the high pass it reshapes the layout, and the fine detail it adds comes from the last steps anyway. Boost the high pass, Detail Daemon on the low pass.
• Two-pass setups: works chained (boosted 0-15 with leftover noise, then a plain sampler 15-25), same as one sampler.
• Off is really off: grow_at_step 0 is pixel-identical to KSampler / KSampler Advanced.

---

Open questions

• Video (Wan, LTX, MiniMax H3) needs a guider-level two-pass version to work with their samplers. Parked for now.
• Has anyone tried a per-model switch point learned from the spectrum, rather than one number per model family?

Feedback welcome, especially from anyone running it on models I have not tested.