H3 images look crunchy? 3 settings (size, sampler, prompt) fix it
Obvious-Bad7548 · reddit · 2026-11-15
A Reddit user shares how to fix "crunchy faces and blocky skin" in generated images—same model, prompt and seed, only three changes:
- Resolution: the model was trained around 16MP; using 3552x4736 for portraits. The default 4MP is where broken teeth come from. Going 4MP→16MP adds only 9 seconds of sampling.
- Sampler: SamplerLCM (built into ComfyUI) with noise 0.9 instead of 1.1; euler at 16MP still gives blocky, oversharpened skin.
- Prompt: short prompts yield a glossy doll face; write long and describe it like a phone photo—camera, light, skin.
Comparison images (4MP vs 16MP vs 16MP+LCM, short vs long prompt), background swaps, product label text, a magazine cover and try-on examples included; workflow files in the first comment.
Related event: Three Settings Fix Blurry H3 Image Outputs(2 posts)→
More from Multimodal
- A watercolor Midjourney SREF code blending Scandinavian minimalism with kawaii — tisch_eins · 2026-11-15
- AskEditor embeds a hand-editable video timeline inside Claude via MCP — MrxAnomynous · 2026-11-15
- ESCAPEMENT: a 3-minute music film coded by Fable 5.2 with no video model — CoolTopicsNow · 2026-11-15
- Dev demos AI-generated AAA-quality game, predicts open-source games boom next year — ChrisGPT · 2026-11-15
- A canopy shot generated with Seedance 2 — Strange-Material-644 · 2026-11-15
- AI Short Film 'The Missing Colors' Reveals Full Stack: Midjourney, Seedance, Suno, ElevenLabs — LudovicCreator · 2026-11-15