PANDATREE commited on
Commit
35d0180
·
verified ·
1 Parent(s): 401ac6d

Link released subject-condition data, training guide, and BBox token-control examples

Browse files
Files changed (1) hide show
  1. flux2-klein-9b/README.md +14 -4
flux2-klein-9b/README.md CHANGED
@@ -61,10 +61,13 @@ background token values into the sub branch; PE routing changes positions.
61
 
62
  ## Data and examples
63
 
64
- The FLUX training code has a subject-driven data path with separate background,
65
- target, subject-reference and subject-crop information. This checkpoint release
66
- does not upload new dataset files or establish a new public subject-dataset ID.
67
- The original Qwen dataset link is not a substitute for that subject-driven data.
 
 
 
68
 
69
  The accompanying GitHub example is an existing 2026-09-04 Gradio comparison,
70
  not a newly generated benchmark. Its metadata identifies these two eval exports
@@ -73,6 +76,13 @@ is **not** evidence for a token-removal ablation. The source Gradio backend uses
73
  a CUDA random generator on CUDA; the historical example must not be described
74
  as CPU-RNG output. No new image generation was run for this release.
75
 
 
 
 
 
 
 
 
76
  ## License
77
 
78
  See [LICENSE.md](LICENSE.md) and [NOTICE.md](NOTICE.md). These weights are FLUX
 
61
 
62
  ## Data and examples
63
 
64
+ The FLUX training data is now released as the
65
+ [subject-condition extension](https://huggingface.co/datasets/PANDATREE/BRIDGE/tree/main/subject_condition)
66
+ of the existing BRIDGE dataset: 27,834 train / 3,092 test rows and 30,926 selected
67
+ subject references generated with `Qwen/Qwen-Image-Edit-2511`. See the
68
+ [full training guide](https://github.com/XIONGPEILIN/BRIDGE/blob/master/flux2/TRAINING.md).
69
+ BRIDGE uses the same method on both backbones: Qwen uses LoRA plus gates, while
70
+ FLUX uses full-transformer plus gate training.
71
 
72
  The accompanying GitHub example is an existing 2026-09-04 Gradio comparison,
73
  not a newly generated benchmark. Its metadata identifies these two eval exports
 
76
  a CUDA random generator on CUDA; the historical example must not be described
77
  as CPU-RNG output. No new image generation was run for this release.
78
 
79
+ Additional [same-BBox-weight cropped-sub comparisons](https://github.com/XIONGPEILIN/BRIDGE/tree/master/flux2/examples/bbox-token-control)
80
+ are now available. Those switch the same BBox checkpoint from dense bbox sub/PE
81
+ support to mask-selected sparse sub/PE support. Six custom-input examples are
82
+ included; internal-dataset examples are excluded. Both token support and PE
83
+ candidate support change. Use the public launcher's `--bbox-protocol-compare`
84
+ option for this mode rather than the default two-trained-checkpoint comparison.
85
+
86
  ## License
87
 
88
  See [LICENSE.md](LICENSE.md) and [NOTICE.md](NOTICE.md). These weights are FLUX