Skip to main content
Back

Nebsh's Compact LoRA Brings Seedream 5.5 Lighting to Open MiniMax H3

A community-trained LoRA adapter pushes MiniMax H3 Image toward Seedream 5.5's cinematic photographic look, with a 1,134-image dataset and 6,000 training steps.

Model, Full Access
  • Seedreamstyle-H3-6K is a photographic style LoRA for MiniMax H3 Image targeting Seedream 5.5's cinematic look.
  • Trained on 1,134 captioned images, rank 16, 6,000 total steps using Automagic3 optimizer at lr 2.8e-4.
  • Use with trigger seedreamstyle., strength 1.0, LCM sampler, 6 steps, CFG 1 in ComfyUI.
  • Comparison gallery shows 20 scenes benchmarked against base H3, 2K checkpoint, and Seedream 5.5 references.
  • Inherits MiniMax H3-Image community license, which excludes US, EU, UK and South Korea.
  • All nine checkpoints from 2K to 6K are published for comparison and experimentation.

Nebsh has published Seedreamstyle-H3-6K, a photographic low-rank adaptation, or LoRA, for MiniMax H3-Image. A LoRA adds a compact set of supplemental weights to a base model. This adapter targets portraits and other human scenes, pushing H3 toward traits associated with ByteDance’s closed Seedream 5.5 model: practical lighting, film grain, stronger contrast, and textured skin.

According to MiniMax’s H3 announcement, the open-weight base model has 33.1 billion parameters and generates images at roughly 16 megapixels, including 4096 by 4096 output. It can produce an image in about six sampling steps, runs without classifier-free guidance, and accepts as many as 16 reference images. Those capabilities make H3 a fast, flexible base for testing whether lightweight tuning can reshape its comparatively neutral rendering of people.

Six steps and one trigger

The supplied ComfyUI workflow layers the adapter over a pruned, 8-bit H3 checkpoint. Prompts must begin with seedreamstyle., and the author recommends a LoRA strength of 1.0. The workflow uses LCM, a sampler suited to low-step generation, with CFG set to 1.

Recommended inference settings