// the find
comfyanonymous/ComfyUI_experiments
Some experimental custom nodes.
A grab-bag of experimental custom nodes for ComfyUI from the ComfyUI author himself — tonemapped CFG, rescale CFG (implementing the arxiv 2305.08891 paper), block-level model merging for SD1.x/SDXL, and a reference-only controlnet node. For ComfyUI power users who want to try specific sampling/merging tricks without waiting for them to land upstream or forking anything.
Each node implements a specific, named technique (not generic 'improvements') — e.g. rescale CFG is a direct port of a published paper's math. Files are single-purpose and self-contained, so you can copy just the one .py you want instead of pulling in a whole package. Block-level model merging gives per-layer merge ratios instead of ComfyUI's single global ratio, which is a real capability gap this fills. Ships an example workflow JSON for reference_only, so there's at least one concrete usage reference beyond the code.
No tests, no CI, and nothing verifying these still work against current ComfyUI internals — last push was May 2024 and ComfyUI's internal node/model APIs change often enough that silent breakage is likely. Documentation is one sentence per file; there's no explanation of what tonemap noise actually looks like, when you'd reach for it over rescale CFG, or what parameter ranges are sane. The class name ModelSamplerTonemapNoiseTest literally has 'Test' in it — this is explicitly throwaway experimental code, not something with any stability guarantee. No license file content described beyond its presence, and no versioning or changelog, so you can't tell if a given node's behavior has silently changed since you last copied it in.