Model designs for image generation
last edited
Model designs for image generation in late 2026 favor one single-stream transformer, few sampling steps and shared latent spaces. Most open designs still denoise VAE latents with flow matching, a lineage from DiT and SD3. Outputs feed the five AI art style families.
Single-stream packed sequences
MiniMax-H3-Image, released in November 2026, is the clearest example. Its prompt, reference images and target latent form one packed token sequence for a single-stream transformer. The design is cut from the omni-modal MiniMax H3 video model, minus its audio layers. Generation and editing share one checkpoint, and the text encoder comes unchanged from the parent model.
Few steps, no guidance
Step distillation now ships on release day. The H3 image design is distilled to six steps without classifier-free guidance. It still renders about 16 megapixels natively, up to 4096 pixels square. A community 2-step LoRA cuts that to two evaluations of its 33.1B transformer. Qwen-Image 2.1 offers turbo files for five to eight steps.
Shared latent spaces
Labs now often reuse one autoencoder instead of training their own. MOSS-Image and Fibo 2.0 both ship the FLUX.2 autoencoder, so latents and tooling carry across models. Representation autoencoders go further, using a frozen pretrained vision encoder as the latent space. Related: weight streaming, Gaussian splatting, 2026 aesthetics and what the SD community still uses.