Repository brief

Comfy-Org/ComfyUI

Read the upstream summary on the left, browse the cached forks below it, and load each fork comparison into the right-hand panel.

Cached analysis

cached 2026-03-29T22:24:06.690Z

Comfy-Org/ComfyUI

ComfyUI is a large, actively maintained Python-based visual AI engine and application for building diffusion pipelines with a graph/nodes interface. It is very widely adopted, with 107,253 stars and 12,377 forks, and it is still receiving frequent commits as of 2026-03-29.

GitHub

Stars107,253

Forks12,377

Default branchmaster

Last pushed2026-03-29T21:11:31Z

Best maintainedpatientx/ComfyUI-Zluda

Closest to upstreamltdrdata/ComfyUI

Most feature-richhiddenswitch/pip-and-uv-installable-ComfyUI

Most opinionatedZHO-ZHO-ZHO/ComfyUI-ZHO-Chinese

Forks

Choose a fork to inspect

6 cached fork briefs

Fork comparison

patientx/ComfyUI-Zluda

active

significant_divergence

Choose this fork if your priority is AMD GPU support and you want a ComfyUI distribution already shaped around ZLUDA and advanced workflows. Stay with upstream if you want the broadest, least specialized, most portable ComfyUI baseline.

Likely purpose

Make ComfyUI run better on AMD hardware through ZLUDA and ship ready-made workflows/nodes for video and image-generation pipelines that target that stack.

Best for

AMD GPU users who want ComfyUI with ZLUDA acceleration.; Power users who prefer a preconfigured workflow library for video and advanced image pipelines.; People willing to trade upstream generality for better hardware-specific performance and specialized nodes.

Additional features

ZLUDA integration for better AMD GPU performance.
AMD-specific runtime and attention/kernel support under `comfy/customzluda` and `comfy/flash_attn_triton_amd`.
Prebuilt workflows for Wan 2.2 image-to-video, text-to-video, self-forcing, LTXVideo 2.3, Flux, and Qwen image editing.
Custom CFZ nodes and conditioning-caching workflows for more specialized pipeline reuse and memory handling.
Fork-specific packaging and setup changes, including Windows batch helpers and workflow-focused distribution assets.
Changes agent workflow behavior, such as prompt handling, handoffs, worktrees, memory, or orchestration flow.

Missing features

Dynamic VRAM is disabled when using the torch compiler, which reduces one upstream memory-management path.
The fork is effectively specialized for AMD/ZLUDA users, so upstream's more general out-of-the-box focus for non-AMD setups is harder to rely on.
Users who want only upstream-stable behavior may need to avoid fork-specific custom nodes/workflows that are not part of mainline ComfyUI.
Some upstream behavior appears to be removed, disabled, or intentionally stripped down in this fork.

Strengths

Strong fit for AMD GPU owners who want ComfyUI to be more usable or faster on their hardware.
Includes many ready-made workflows instead of forcing users to build advanced graphs from scratch.
Still tracks upstream closely enough to remain current, while adding a large amount of fork-specific functionality.
Appears tuned for practical production workflows around video generation and caching/memory management.

Risks

Large fork delta means behavior may diverge from upstream and future merges may be harder.
AMD/ZLUDA-specific code increases maintenance burden and narrows the supported hardware path.
Extra custom workflows and nodes can make the environment more complex to reproduce or debug.