A reproducibly generated mixed-FP8 E4M3FN quantized derivative of Alibaba-PAI's
FLUX.2-dev-Fun-Controlnet-Union-2602.safetensors, produced independently by the JLC project.
This release is intended for the FLUX.2-dev Union ControlNet path in ComfyUI. It is not a
FLUX.2 Klein ControlNet.
License: FLUX Non-Commercial License v2.1. Non-commercial use only. See
LICENSE and NOTICE.md.
This is an independent derivative release. It is not an official product of, and has not
been endorsed, approved, or validated by, Black Forest Labs Inc. or Alibaba-PAI.
The published checkpoint is byte-identical to the authoritative reproducibly generated and
validated release artifact.
Examples
The examples below were generated with the release checkpoint. Each uses a single
preprocessed ControlNet hint at strength 0.50, active across the full 0.00–1.00 inference
interval. Only the hint / generated-output pair is shown; the licensed Adobe Stock source
photographs are not embedded in these example files.
The selected set intentionally spans several different conditioning representations while
showing two challenging articulated subjects.
DWPose — Ballerina
Pose-skeleton conditioning.
DWPose ballerina hint and generated output
HED — Skater
Edge / line-structure conditioning.
HED skater hint and generated output
DepthAnythingV2 — Skater
Depth / volume conditioning.
DepthAnythingV2 skater hint and generated output
DSINE Normal Map — Ballerina
Surface-orientation / normal-map conditioning.
DSINE normal-map ballerina hint and generated output
Color — Ballerina
Low-frequency color / spatial-layout conditioning.
Color ballerina hint and generated output
ComfyUI workflow
This sample ComfyUI workflow demonstrates multi-control conditioning with the JLC FP8 checkpoint. It combines DWPose, HED, and DepthAnythingV2 hints and can be used as a starting point for reproducing examples like those shown above. Both the PNG and JSON are ready for direct loading into ComfyUI; the PNG can also be drag-and-dropped onto the ComfyUI canvas.
This is a selective mixed-precision checkpoint, not a global FP8 cast.
The JLC quantization policy stores 48 large weight tensors as FP8 E4M3FN, retains 28 protected
tensors in BF16, and adds 96 FP32 scale tensors used by the mixed-precision execution path.
Source SHA-256 fingerprint:516532a8…b0a7a794 — full checksum
Validation
The release artifact passed the following acceptance checks:
toolkit-driven quantization: PASS
independent reopened-file verification: PASS
172 / 172 tensor payloads reproduced byte-for-byte against the independently generated FP8 validation reference
native ComfyUI mixed-precision load: PASS
matched runtime validation at 1024 × 1536: pixel-identical against the validated FP8 reference artifact
These results establish deterministic reproduction and successful execution in the validated environment. They do not claim universal numerical identity across every hardware platform, workflow, attention backend, resolution, or future software version.
ComfyUI use
This checkpoint is supported through JLC Flux2 ControlNet v1.1.0 or later:
Place the checkpoint in your ComfyUI ControlNet model directory and select it with JLC Flux2 ControlNet Loader.
The unified loader automatically supports compatible dense BF16 FLUX.2 Fun ControlNet checkpoints and the JLC mixed FP8/BF16 Union-2602 checkpoint through the same runtime node family. No separate FP8 loader, Apply node, or FP8-specific workflow path is required.
The remaining JLC Flux2 ControlNet nodes—including Apply, Orchestrator, reference-image conditioning, cache preparation, and experimental in/out-paint support—operate through the same shared ControlNet runtime interface.
Validated example workflows are provided with the JLC Flux2 ControlNet release.
Reproducibility
The quantization/calibration toolkit and detailed reproducibility, validation, provenance, and methodology materials used to produce this checkpoint are currently being prepared for public release.
The release artifact itself has already completed the validation summarized above.
The toolkit publication will provide the supporting material for reproducing and independently verifying the mixed FP8/BF16 checkpoint from the original BF16 source. Model inference does not depend on the toolkit and is supported through the public JLC Flux2 ControlNet package linked above.
License and attribution
The checkpoint is a derivative of a FLUX model and is not relicensed under MIT,
Apache-2.0, GPL, or another permissive software/model license.
The model derivative is distributed under the FLUX Non-Commercial License v2.1
included in LICENSE. No commercial or production rights are granted by this repository.
The required Black Forest Labs attribution, source-model attribution, modification statement,
and non-endorsement statement are reproduced in NOTICE.md.
The JLC software used to produce and run the model is licensed separately in its respective
software repositories; those software licenses do not relicense this checkpoint.
Author / derivative project
Independent mixed-FP8 quantization and validation work:
JLC / José Luis Cordova
The repository metadata points to Black Forest Labs' current FLUX.2-dev license page; the complete license text is also included locally in this repository.
