LCM img2img Sampler

The pack's README is blunt about this one: "for basic img2img, you can just use the LCM_img2img_Sampler node." It's the same idea as the plain LCM Sampler - a self-contained node running the diffusers LCM pipeline in 4-8 steps instead of the usual 20-50 - except it starts from an image you feed it instead of pure noise. Feed it a source image and a prompt, and it pushes the image toward the prompt in a handful of fast steps.

This whole pack, worth saying up front, predates ComfyUI's native LCM scheduler support. The README carries an archival notice pointing at the official implementation now that it exists in core ComfyUI. This node still works and is genuinely the simplest way to get this specific behavior - one checkpoint, one pipeline, nothing else to configure - but don't expect it to be the state of the art for img2img speed techniques in 2026.

How it works

Same mechanism as the txt2img version, with an extra step at the front: your input image gets noised up to a degree controlled by prompt_strength, then the LCM_Dreamshaper_v7 pipeline denoises it back down in 4-8 steps, guided by your prompt. It's the LCM equivalent of the denoise-strength slider you'd see on a normal img2img workflow - low values stay close to your source image, high values give the prompt more room to override it. Like the txt2img sampler, this one loads its own checkpoint internally via diffusers, so there's no MODEL or VAE to wire in.

The inputs and outputs that matter

images is your source image. prompt_strength (0-1, default 0.5) is the one setting that actually changes the character of the result - nudge it down for a subtle edit that keeps your composition, up toward 1 to let the prompt take over more. positive_prompt is plain text, same as the txt2img node - no negative prompt slot here either. The rest matches the base sampler: seed, steps (default 4, built for 4-8, don't push it past that expecting more detail), cfg (default 8 in the widget, but LCM wants it down around 1-2 since the guidance is already baked in from distillation - leave it at 8 and you'll get harsh, oversaturated results), width/height (512-768, SD1.5-class), num_images for batching several variations at once, and use_fp16 for the low VRAM footprint. Output is a finished IMAGE.

If you're after vid2vid rather than a single still, this node is the building block but not the whole answer - the README points to a separate pack, ComfyUI-VideoHelperSuite, for the Load Video/Video Combine nodes that feed frames through this sampler one at a time.

Installing it

git clone https://github.com/0xbitches/ComfyUI-LCM into custom_nodes/, restart ComfyUI, or search "Latent Consistency Model for ComfyUI" in ComfyUI Manager. No manual model download - LCM_Dreamshaper_v7 auto-downloads into your HuggingFace cache on first run, so the first generation pauses to pull a few gigabytes.

Common issues

The README's documented fix is for a tokenizer error, Non-consecutive added token '<|startoftext|>', caused by a mismatched cached tokenizer - fixed by hand-editing added_tokens.json in the cache folder (exact replacement is in the README). Beyond that, several people in the original release thread hit FileNotFoundError/PermissionError around the HuggingFace cache path on first download, mostly on Windows - deleting the specific models--SimianLuo--LCM_Dreamshaper_v7 folder from the cache and letting it redownload usually clears it. And because this pack talks to diffusers directly instead of reusing ComfyUI's own model stack, watch for the standard shared-environment dependency conflicts if you've got other diffusers-based nodes installed - everything lives in one Python environment here.