Emu 3.5 Official T2I (Auto Size)

Emu 3.5 Official T2I is the pack's "do it the way BAAI wrote it" node. While the V2 samplers add convenience knobs and tiled decoding, this one deliberately mirrors the official Emu3.5 example_config_t2i.py structure as closely as the pack author could manage - same task template, same logits processor, same auto-size behavior. If you want a baseline of what the model itself does with no cleverness layered on, this is it.

The "Auto Size" in the display name is the point. You pick an aspect ratio from the dropdown and the node derives the token-grid dimensions itself (a 1:1 image is 64x64 latents = 1024x1024 pixels; a 16:9 goes to 752x1360, and so on). Emu 3.5 generates images as a stream of visual tokens - between <|image start|> and <|image end|> markers - rather than denoising latents, so the node's real work is building the correct USER: {prompt} ASSISTANT: chat template, feeding unconditional logits through the CFG logits processor, and then decoding the token stream back through the vision tokenizer (the IBQ VQ-VAE with its 262,144-entry codebook).

The inputs that matter

• prompt - plain English works; this model likes detailed, natural-language descriptions.
• aspect_ratio - default, 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3, 21:9. default lets the model pick.
• cfg_scale - guidance, default 5.0. Emu's CFG works differently from diffusion CFG (it's applied at the logits level), but the intuition is the same: higher = stricter prompt adherence, and values that are too high can smear quality.
• seed - fixed seed for reproducible outputs, 0 for random.

Outputs

• image (IMAGE) - the generated image, ready for a preview or Save Image node.
• text (STRING) - any text the model emitted alongside the image (it can narrate before drawing).
• reasoning (STRING) - chain-of-thought output when the model produces it.

Install

It's in the pack with the rest:

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

Weights in ComfyUI/models/emu35/: BAAI/Emu3.5-Image (34B) plus the vision tokenizer, or the NF4 quantized build for 24GB cards.

Who should reach for it

Honestly? Mostly people who want the unvarnished reference behavior or are chasing a bug in the fancier samplers. The V2 T2I sampler is the better daily driver - it adds separate text and image top_k / temperature controls and a tiled decode path that keeps big images from OOMing on smaller cards. But there's real value in having a known-good baseline: if your V2 output looks wrong, run the same prompt through Official T2I to see whether the model itself is the problem or the extra knobs are. And brace for the wait - autoregressive generation means a 1024x1024 is roughly 4,000 tokens at 5-ish tok/s on a 96GB card, so about twelve minutes a pop.