52 lines
2.3 KiB
Markdown
52 lines
2.3 KiB
Markdown
# Hugging Face Community Notes
|
|
|
|
These are useful `FLUX.2-klein` community findings from Hugging Face pages that did not all ship clean standalone workflow JSON files, but still contain setup or optimization advice worth keeping.
|
|
|
|
## Unsloth GGUF discussion
|
|
|
|
Source: <https://huggingface.co/unsloth/FLUX.2-klein-base-9B-GGUF/discussions/1>
|
|
|
|
What looked useful:
|
|
|
|
- At least one user reports a working `9B base` path by starting from the official Comfy `flux2 klein base 9b` workflow and swapping the diffusion and clip loaders to GGUF-compatible loaders.
|
|
- The discussion explicitly calls out the text-encoder mismatch trap:
|
|
- `4B` uses `Qwen3 4B`
|
|
- `9B` uses `Qwen3 8B`
|
|
- The thread also suggests keeping ComfyUI very current; some of the reported matrix-shape errors appear to come from version mismatch or wrong encoder pairing.
|
|
|
|
Our take:
|
|
|
|
- Good sign that community users are already adapting the official workflows instead of reinventing everything from scratch.
|
|
- If you want a `9B GGUF` branch later, this discussion is one of the better breadcrumbs for why it fails and how people are fixing it.
|
|
|
|
## GuangyuanSD Blitz notes
|
|
|
|
Source: <https://huggingface.co/GuangyuanSD/FLUX.2-klein-9B-Blitz-ComfyUI>
|
|
|
|
What looked useful:
|
|
|
|
- The page is strongly oriented around low-step speed-first `9B` usage.
|
|
- Recommended settings are very explicit:
|
|
- `4-5` steps
|
|
- `CFG = 1`
|
|
- The author says their environment was `ComfyUI 0.11`.
|
|
- The repo also mentions LoRA-adapter style application paths and Blackwell-oriented quantization variants.
|
|
|
|
Our take:
|
|
|
|
- Even though this page did not expose a clean workflow JSON in the fetched content, the settings are concrete enough to matter.
|
|
- If you try community `9B Blitz` variants later, these are the first settings I would mirror instead of guessing.
|
|
|
|
## ponpoke uncensored encoder
|
|
|
|
Source: <https://huggingface.co/ponpoke/flux2-klein-9b-uncensored-text-encoder>
|
|
|
|
What looked useful:
|
|
|
|
- It is not a workflow, but it is directly usable in ComfyUI as a replacement text encoder placed in the `models/clip` path.
|
|
- The author explicitly says to pair it with the official `FLUX.2-klein-9B` DiT and standard or GGUF-compatible loader nodes.
|
|
|
|
Our take:
|
|
|
|
- Not something to treat as a default, but worth remembering if you ever want alternate text-encoder behavior in a `9B` setup.
|