[TroyHernandez]

Adds FLUX 3 Image Klein (black-forest-labs/flux-3-image-klein): text-to-image and single-reference editing, ported from the diffusers pipeline. Version 0.2.2.16.

download_flux3_klein()   # gated: accept the FLUX Kommunity License + HF_TOKEN; ~20 GB, one-time fp8 quantize to 5.3 GB
pipe <- flux3_load_pipeline()
txt2img_flux3(caption, pipeline = pipe, seed = 7)
img2img_flux3(caption, image = "photo.png", pipeline = pipe, seed = 7)

What's in it

• flux3_transformer(): per-mode blocks (image, reference, text; 5 each) then 28 joint blocks, per-token timestep modulation, target-timestep embedder. Built as a skeleton so a bf16 load peaks at ~15 GB host RAM instead of 22 GB.
• Qwen3-VL-4B text model through load_qwen3_text_encoder() (layers 4..32 stacked, prompt padded to a multiple of 80).
• flux3_preprocess_image(): reference kept whole, scaled to fit max_area (default 4 MP), reflect-padded to the 16 px grid, output cropped back.
• fp8 quantization (cast census 265) and the gated download, pinned to revision 3373376.
• diffuseR takes the caption as given. BFL recommends a prompt upsampler, and expanding short prompts is the caller's job.

Shared fixes

• VAE mid-block attention now uses fused SDPA. A 2048x2048 image allocated an 8 GiB score matrix and ran out of memory on 16 GB. This code is shared by the SD, SDXL, FLUX.1, FLUX.2 and FLUX 3 VAEs.
• The Qwen3 encoder is built at its target dtype, which removes a 16 GB fp32 intermediate when loading in bf16. Requesting the last layer now returns the post-norm state, matching transformers.
• flux2_feed_forward uses linear_noinit.

Validation (real weights)

• Tokens and prompt template are identical to the shipped processor. Qwen3-VL embeddings are bit-identical at fp32.
• Transformer at fp32, depth-reduced but covering every block type: max abs error 2.9e-6, with and without a reference.
• 1 MP edit on the GPU against diffusers, each with its own text encoder: R bf16 52.2 dB PSNR, R fp8 36.7 dB.
• RTX 5060 Ti 16 GB, fp8:
  • 1024² text-to-image: 44 s, peak 11.7 GB allocated.
  • 1 MP edit: 50 s, peak 13.8 GB allocated.
  • 4 MP edit: 285 s, peak 14.1 GB allocated.
• tinytest: 1540 passed, 0 failed (53 new).

Not in this PR

• An 8 GB mode (fp8 text encoder plus a block-streamed transformer).
• Multiple references.
• Fleet entries.

The weights are under the FLUX Kommunity License and are never redistributed; see inst/COPYRIGHTS.