An 8B single-stream DiT that Baidu open-sourced, and the interesting part is not general image quality but what it does with text. Long prompts, dense Chinese copy and structured layouts come out closer to the brief than two established rivals — while multiple people in one frame start sprouting extra hands.

Evidence note: this episode ships no workflow JSON, so no node counts or sampler values are claimed. Timings and pixel figures below are the author's spoken measurements.

When this is useful

  • Posters, packaging and social cards where the copy has to be legible and correctly placed.
  • Long, structured prompts with several objects that all need to appear.
  • Chinese-language infographics and classroom-style diagrams.
  • Fast drafts: a turbo variant reports roughly three seconds per 1024x1024 image after the model is loaded.

How this differs from the previous build

Where it differs is architectural and behavioural.

  • Single-stream DiT, the same family as the other Turbo model it is compared against, with both accelerated and non-accelerated builds available.
  • Prompt polishing is on by default in the official text-to-image graph — one boolean turns it off.
  • Text is treated as a first-class citizen. The comparison tests are all text-heavy: a blackboard maths problem, a table, a poster, a product label.
  • Realism skews processed. The author's read is that models which over-render portraits tend to look most convincing on objects, and that holds here.

Running it

  1. Deploy the mirror, open a terminal in JupyterLab, paste the startup command, then open ComfyUI from the panel.
  2. Pick the Turbo text-to-image graph for the default path; it ships with prompt polishing enabled.
  3. Set width and height by megapixel budget, then set the ratio.
  4. Enter the prompt. For a fair comparison, disable polishing via the boolean and run again.
  5. Expect a slower first run while the model loads; subsequent 1024x1024 images land in roughly three seconds.
  6. For image-to-image, add the source image, set size and denoise, and run.
  7. For tiled upscale, load the image, set the target — eight million pixels in the demo — and run.

Honest limits

  • Multiple people in one frame produce extra hands and feet — the author's clearest failure case, and a direct counterweight to its text strengths.
  • Portrait quality beats the Qwen comparison slightly but sits below the Z-Image comparison.
  • Aesthetics is not the selling point; the author says so directly.
  • The comparisons use an eight-step distilled variant throughout, so non-distilled behaviour is untested.
  • No workflow JSON ships, so nothing here is a reproducible parameter set.

Episode resources

Mirror (Compshare, referral code preserved)