ELLA Sampler

This is the node that actually generates, in the ELLA half of this pack. ELLA Sampler takes the reconstructed pipeline from ELLA Model Loader, the pre-computed T5 embeddings from ELLA T5 Embeds, and runs them through diffusers to produce your image. Everything before it is setup; this is where pixels come out.

The important detail is what it does not take: a prompt text box. Unlike a normal KSampler, the prompt here is invisible because it's already been encoded upstream by ella_t5_embeds. The sampler calls the diffusers pipeline with prompt_embeds and negative_prompt_embeds directly, which is also why there's a gnarly workaround buried in the code: diffusers concatenates positive and negative embeddings way too early, but ELLA's resampler is time-aware and needs each sample's true token length. The wrapper records those lengths and trims per-sample inside the UNet proxy so ELLA sees the real token count, not a padded one. It works; you just can't cheat and feed it a plain CLIP embedding.

Inputs that matter:

• ella_model (ELLAMODEL) and ella_embeds (ELLAEMBEDS) - both required, both from their named loader/encoder nodes.
• steps - 1–200, default 25. guidance_scale - default 10. seed - standard, and note the batch size isn't set here: it comes from the embeds, and the sampler generates that many images with seeds seed through seed+batch_size-1.
• scheduler - same ten-choice dropdown as the diffusers sampler; the DPMSolverMultistepScheduler default is right for most checkpoints.
• width / height - and here's the fun one: up to 8192, four times the 2048 cap on the non-ELLA sampler. High resolution is the point, because the bundled hidiffusion toggle (optional, off by default) applies the same-lab HiDiffusion patch that keeps SD 1.5 coherent at big sizes. At 512 you don't need it; at 1024+ that's what it's for.

The single output is images (IMAGE), which feeds PreviewImage or a save node like anything else.

Where people get burned: expecting KSampler-style behavior and getting confusions instead. There's no negative-prompt text field here - the negative is whatever ella_t5_embeds produced (an empty string by default). And because this is a diffusers pipeline rather than ComfyUI's native sampler, your usual sampler_1/seed_node tricks don't apply; everything lives in this node's inputs. Since the pack is deprecated in favor of Tencent's official ComfyUI-ELLA nodes, treat this as "use if it's already installed," not "install this new thing." It still works, but it's frozen at May 2024.