Diffusers Sampler

This node is the control group in ELLA's experiment. The whole point of kijai's ComfyUI-ELLA-wrapper pack is putting Tencent's ELLA text encoder on top of an SD 1.5 model, and this sampler is the "without ELLA" side of that comparison. Feed it a DIFFUSERSMODEL from Diffusers Model Loader and it runs a plain diffusers StableDiffusionPipeline with the ordinary CLIP text encoder - no T5, no 6 GB download, none of the ELLA machinery. Same checkpoint, same seed, and you can finally see what ELLA actually buys you.

Mechanically it's a full diffusers pipeline in a node. The prompt and negative prompt get encoded by the stock CLIP text encoder, the scheduler is swapped in at runtime from a ten-choice dropdown, and the whole thing runs under the same autocast/dtype handling ComfyUI sets for your GPU. It returns ordinary IMAGE tensors straight out of PreviewImage or a save node.

The inputs you'll actually touch:

• prompt / n_prompt - multiline strings. Defaults are literally "positive" and "negative"; replace them with your real prompt.
• steps - 1–200, default 25.
• guidance_scale - default 10, which is the SD 1.5 sweet spot.
• seed - standard.
• scheduler - the default DPMSolverMultistepScheduler is fine; you'd only switch for LCM or TCD distilled LoRAs.
• width / height - 64–2048 in steps of 64. The hidden gem here is hidiffusion: flip it on and the node applies the bundled HiDiffusion patch (same lab as ELLA) that lets SD 1.5 render at 1024+ without the classic duplicate-object meltdown.

The one output, images, wires into anything that takes an IMAGE.

The honest use case is A/B testing. Wire up CheckpointLoaderSimple → Diffusers Model Loader → Diffusers Sampler as your baseline, then build the ella_model_loader + ella_t5_embeds + ella_sampler path beside it, keep the seed fixed, and compare. Community testing back in 2024 found ELLA clearly beats CLIP on prompt adherence but doesn't quite reach a properly fine-tuned SDXL model, so a baseline isn't just nice-to-have - it's how you decide whether the extra GBs are worth it. If you decide they're not, this sampler is also just a perfectly serviceable SD 1.5 diffusers sampler on its own.

One warning up front: this pack is effectively retired. The README starts by pointing you to Tencent's official ComfyUI-ELLA nodes instead, and the last commit was May 2024. It still runs - kijai's wrapper was the one people reported "just worked" - but don't expect updates, and check the official pack first if you're starting fresh.