1,000 Same-Seed Renders Later: How to Actually Tune Qwen-Image 2.1

Glad_Captain8190 · reddit · 2026-10-18

The author ran 1,000 Qwen-Image 2.1 renders in ComfyUI on a 16 GB card, same seeds throughout.

Guidance

Official template is CFG 1/25 steps. CFG 3 + negative prompt + APG and FreSca patches keeps composition (same seed/size) but restores what CFG 1 flattens: embroidery stitches, rivets, pupils, contrast
Cost: a second pass per step — 4MP/20 steps goes from 120-135s to 255-275s; the slower setup won all 10 test pairs. Raising CFG 1 to 30-40 steps restores sharpness but not materials

Size/steps change the picture, not its resolution

Same seed at 1MP vs 4MP yields different images; composition held in 8/10 pairs between 2.5-4MP, never from 1MP. 20 vs 25 steps are nearly identical. Pick size first, then hunt seeds: 2.5MP/20 steps for web, 4MP for prints

Turbo and Guided Turbo

Turbo checkpoint (8 steps, CFG 1): 9-10s at 1MP, 38-39s at 4MP — 7x faster but drifts to photorealism (embroidery becomes a crocodile photo). The LoRA version is slower; download the checkpoint
Surprise: keep Turbo's checkpoint and 8 sigmas but set CFGGuider to 3, add negative prompt + APG + FreSca — 66-71s at 4MP brings back detail and saturation. 20 steps with the simple scheduler overcooks it

Traps

The 2025 Lightning LoRA won't load on 2.1 (3072 vs 4096 attention width) and silently renders the bare model
Invented lettering on shopfronts unless prompted "no lettering"; unknown characters get swapped for lookalikes

Full node graph, negative prompt and prompt structure at the linked guide.

More from Multimodal