Emu 3.5 Interleaved (Story/HowTo)

This is the node that shows off what Emu 3.5 can do that diffusion models can't. Emu 3.5 Interleaved generates a full piece of writing with images embedded in it - a story where the model writes a paragraph, decides it's a good moment for an illustration, generates the image, and keeps going. One generation, text and pictures interleaved, because in Emu's unified token space text tokens and visual tokens are just different entries in the same vocabulary. The base model was literally trained to do this.

You pick a task_type - story or howto - and the node runs with task-specific sampling presets until it has written max_images images' worth of content (default 4, up to 10), using a stopping criterion that counts only generated image-end tokens. The output gives you the full text and the images separately: full_text is the prose with the image slots in place, and images is the batch of illustrations in order.

The inputs that matter

• prompt - the topic or story seed. "Tell me a story about a robot learning to paint" is the default for good reason.
• task_type - story or howto. HowTo mode is the sleeper hit - step-by-step tutorials with a diagram per step.
• max_images - 1 to 10, default 4. Each image is thousands of tokens; this is your runtime dial.
• cfg_scale - default 3.0, lower than T2I's 5.0; interleaved generation gets brittle at high guidance.
• seed - fixed for reproducibility.
• reference_image - optional; gives the model visual context to build on.

Outputs

• images (IMAGE, a list) - the illustrations, in generation order.
• full_text (STRING) - the complete story/tutorial text.
• reasoning (STRING) - chain-of-thought, when present.

The catch, and it's a big one

This node needs the base Emu 3.5 model (65B parameters), not the 34B Emu3.5-Image model - the code even prints a warning if it detects you've loaded the Image model, because the Image model is fine for T2I/X2I but wasn't trained for interleaved output. The 65B base does not fit on a single 24GB card, quantized or not. The community consensus is blunt: story mode is a two-GPU proposition (two 80GB cards for comfortable BF16, or a two-24GB NF4 split via Loader V2's device=auto), and even then you're looking at a long rollout for a multi-image story.

Install

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/EricRollei/Emu35-Comfyui-Nodes.git emu35
cd emu35
pip install -r requirements.txt

Weights: huggingface-cli download BAAI/Emu3.5 --local-dir models/emu35/Emu3.5 (the base - it's a big download) plus the vision tokenizer. If you're on a single consumer card, honestly consider whether you want this node yet: the story feature is the pack's most impressive demo and its least practical one without the hardware.