Image models split: a 5.6B editor for your GPU, flagships for the niche

The tiny editor

Black Forest Labs released FLUX 3 Image Klein, a 5.6 billion-parameter open-weights model for text-to-image generation, image editing, and multi-reference generation with up to 10 reference images. The company's internal evaluations put it first among open-weights models on text-to-image, multi-reference generation, and image editing — and, the company claims, above closed models in the comparison, naming Ideogram V4.5 and FLUX.2 [pro]. The weights ship under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5 million. m:…0525

The story is really about the hardware envelope. The INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM through ComfyUI or Diffusers, and the company says the distilled model generates a 1-megapixel image in about 4 seconds on an RTX 5090. The release includes a full undistilled base checkpoint for fine-tuning and research alongside the distilled low-latency variant. For API users, the company prices the distilled model starting at $0.009 per 1,000 images. Note the sourcing: the rankings and latency figures are the company's own claims — the checkpoints were red-teamed internally and evaluated by external safety partner Cinder — so treat them as vendor numbers until independent benchmarks land. m:…0525

In the workflow: ComfyUI

ComfyUI's integration notes make the practical case. The model runs natively in ComfyUI with bounding-box composition — mark regions on the canvas to set layout — plus multi-reference generation and editing, and, in the project's words, "Generated at 4K, not upscaled into it." Pairing INT8 ConvRot quantization with the step-distilled model makes 1024x1024 generation 4.3x to 9.3x faster than the BF16 50-step baseline: on an RTX 5070, an image that took 71 seconds now takes under 8; on an RTX 4070, 78 seconds drops to under 9. A separate prompt-upsampler model expands short prompts before generation, and the weights are downloadable from Hugging Face. m:…9066

The flagships

At the other end of the market, fal's John Ozuysal published a head-to-head of the two flagship editing models on the fal platform. Nano Banana 2.1 is controlled through a resolution tier and a thinking level — with web search grounding, video and PDF context, a system prompt, and aspect ratios out to 8:1. "GPT Image 2.5 comes in Flare and Sunburst variants controlled through exact pixel size and a quality setting" — with transparent backgrounds, mask_url and 16 references per edit. m:…3101

The niches are complementary, and so is the pricing. Nano Banana 2.1 carries web search, video_url, and pdf_url context that GPT Image 2.5 lacks; GPT Image 2.5 has transparent-background support and reusable mask editing that Nano Banana 2.1 lacks natively. On per-image pricing, Nano Banana 2.1 costs $0.0593 for a 16:9 2K image from a short prompt, while GPT Image 2.5 at 2560x1440 on high quality costs $0.05529 for a 16:9 image. Both are partner models on fal with commercial use enabled on all endpoints. m:…3101

The split is the story — and that reading is the author's synthesis, not either vendor's claim. One lane is shrinking the model until it fits on the GPU a developer already owns: 5.6B parameters, 8GB of VRAM, weights on Hugging Face. The other lane is specialization at the API: two flagships that don't compete on the same editing tasks so much as divide them. For developers, the decision isn't which model is 'best' anymore; it's whether the job is local, cheap, and private, or niche, composited, and billed per use. m:…0525 m:…3101