DiffSynth: Boogu Image Inference

Boogu Image is the one to reach for when your prompt has words in it that need to come out as words. A community test of the Turbo checkpoint asked for a four-pane storyboard with quoted speech bubbles and got readable dialogue back in all four panels - a genuinely rare result - and the model's other strength is public characters and named styles rendered without much fuss. Reception elsewhere was split, mostly on the "looks like AI art" axis, which is a taste argument rather than a capability one.

Required sockets: pipe (from the Boogu Loader), prompt, seed, num_inference_steps, height, width. One output: image → Save Image.

Steps are the whole story

num_inference_steps defaults to 20, but the checkpoint decides: the Turbo template runs 4, the Base and Edit templates run 50 and 50. Run Turbo at 4 and you get a clean image in seconds; run it at 20 and you're burning time on a model that has nothing left to denoise. Base is slower and more variable between seeds - which is what one early tester flagged with Turbo, "images look too synthetic and lack variability between seeds".

Other settings worth knowing:

• cfg_scale - default 4. Reasonable place to start on Base. On Turbo, treat guidance like a distilled model's: low, not high.
• negative_prompt - works while you're at real CFG. At CFG 1 it's discarded entirely.
• edit_image - the second image input, and the reason the Edit checkpoint exists. Wire an image here (plus input_image where the template does) and the node behaves like an instruction editor rather than a text-to-image sampler.
• input_image + denoising_strength - the img2img path on the non-edit checkpoints. denoising_strength default is 1.0, i.e. "ignore the input". Drop it to 0.4-0.6 for a partial re-render.
• sigmas - free text, default empty. A place to override the sampling schedule directly; leave it blank unless you know the numbers you want.
• max_sequence_length (1280) - prompt token budget. Long structured prompts get truncated past this.

Then four size budgets, which look intimidating and are actually your OOM levers, because Boogu has two image pipelines and each has its own limits:

• max_input_image_pixels (4194304 = 4MP) and max_input_image_side_length (4096) - for the main model.
• max_vlm_input_pil_pixels (147456, ~0.15MP) and max_vlm_input_pil_side_length (768) - for the vision-language front end that reads your reference image.

If you're feeding a big photo into the Edit path and the run dies, lower the VLM pair first. It doesn't need full resolution to understand the scene.

Install

# Manager: search "DiffSynth-ComfyUI" → Install → restart
pip install git+https://github.com/modelscope/DiffSynth-Studio.git
cd ComfyUI/custom_nodes && git clone https://github.com/modelscope/DiffSynth-ComfyUI.git

Then open the Template tab and load any DiffSynth Boogu workflow. That gets you the three ModelConfig nodes, the mllm/ processor_config, the Loader and this node in the right order; change the prompt and run.

Troubleshooting

If output is mushy at 4 steps, you're on Base or Edit weights, not Turbo. If the composition ignores your layout instructions, remember Boogu's strength is prompt-level text and subject knowledge, not a JSON bounding-box interface - for precise panel geometry you're fighting the model. Expect the "AI look" to be a real criticism rather than a setup error: this one was never going to be mistaken for a photo. And the first queue on a fresh install just downloads weights from ModelScope; there's no progress bar worth watching, so give it time before deciding it hung.