ELLA T5 Embeds

This node is the whole reason ELLA exists. SD 1.5's weak point was always its 77-token CLIP text encoder - you'd write a gorgeous five-element scene and get three of them, plus artifacts bleeding into the image. ELLA T5 Embeds swaps CLIP for a full flan-t5-xl encoder (T5-XL), and that's where the "semantic alignment" comes from: an LLM-sized text encoder that understands long, compositional prompts instead of mangling them. It's the direct fix for the prompt-adherence complaints that defined the SD 1.5 era.

Mechanically, this node is a text encoder, full stop. You give it text, it runs the T5-XL encoder and returns an ELLAEMBEDS bundle - prompt_embeds, negative_prompt_embeds, and a batch size - that ELLA Sampler consumes. There is no CLIP here, no KSampler, no image. And there's no negative prompt text box either, which surprises people: the negative side is just an empty string, padded per the settings below. ELLA was trained against the empty string as the unconditional, so you mostly don't get to steer the negative.

The inputs that matter:

• prompt - multiline, and it has a long multi-clause demo default so you immediately see the point. Notice that T5 handles it without token bloat; max length is 512 tokens vs CLIP's 77.
• max_length - default 128, up to 512. Only relevant when flexible_max_length is off.
• flexible_max_length (optional, default on) - this is the setting that actually matters. ELLA's resampler is time-aware and wants each prompt at its natural token count, so when this is on, max_length is ignored and every prompt gets its true length. Leave it on.
• fixed_negative - whether the empty-string negative gets padded to max_length (on) or kept at its natural length (off). Default off is correct.
• batch_size - 1–256. This becomes the number of images the sampler generates.

Output: ella_embeds (ELLAEMBEDS), wired into ELLA Sampler.

The cost, and it's real: on first run this node downloads a ~6 GB flan-t5-xl encoder (Kijai/flan-t5-xl-encoder-only-bf16 into ComfyUI/models/t5_model/; older installs use the sharded ybelkada repo). That's a big download and then a few GB of VRAM while encoding - though it offloads back to CPU when done, which is why this pipeline stays livable on 8–12 GB cards.

Install is the pack-standard route: ComfyUI Manager (search "ComfyUI-ELLA-wrapper"), or

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-ELLA-wrapper
cd ComfyUI-ELLA-wrapper
pip install -r requirements.txt

On the portable Windows build, run python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-ELLA-wrapper\requirements.txt from the portable root. Requirements are diffusers>=0.26.0, accelerate, omegaconf, sentencepiece - if your ComfyUI is behind, update diffusers first or the whole pack refuses to import.

Troubleshooting: if you hit "layerNormKernelImpl" not implemented for 'Half' while this encodes - the signature ELLA-family error on fp16 systems - relaunch ComfyUI with --fp32-text-enc added to your launcher args. And remember: SD 1.5 only, and the pack is deprecated in favor of Tencent's official ComfyUI-ELLA. This one works, but it's a snapshot from mid-2024.