Designing a 16-image reference experiment for MiniMax H3 Image room editing
AIdianjiaoshi · reddit · 2026-11-14
A Redditor shared their experimental design for testing how many reference images MiniMax H3 Image actually uses in multi-image room editing, and asked how to quantify drift in unedited regions.
- ComfyUI edit template takes up to 16 inputs; Picture 1 is the base. A 2688×1536 (4MP) canvas assigns furniture, materials, and lighting by number
- Three reference tiers: single, few, and full references; local edits change one object and check whether unspecified regions change
- Resolution fixed near 4MP first; native support is 16MP (4096×4096 ≈ 16.78MP, 4K UHD ≈ 8.29MP)
- The model is 33B params, single-stream DiT without an audio branch, with INT8 diffusion weights, text encoder, and image VAE; 6-step sampling, no CFG
- They plan to log omissions, duplicates, and cross-object bleeding while pinning prompt-expansion toggle, model files, and template versions
More from Multimodal
- Prediction: Pixar-quality animation will explode, but not from video models — ChrisGPT · 2026-11-14
- Negative momentum cuts sampler error 11-21% vs dpmpp_2m at same step count — YakovSava · 2026-11-14
- "Unbelievable" AI video from creator tupacabra stuns the AI community — LinusEkenstam · 2026-11-14
- Redditor shares optimized ComfyUI workflow for MiniMax H3 Image on a 24GB RTX 3090 — Zetalvx-AI-Lab · 2026-11-14
- 20 gorgeous Western-style paintings (likely AI-generated) going viral — xiaosun86 · 2026-11-14
- Kitten TTS 2 ships three tiny open-source TTS models, smallest just 8MB on CPU — pmttyji · 2026-11-14