Welcome FLUX.3 Image Klein in Diffusers

Black Forest Labs released FLUX.3 Image Klein, an open image generation model that runs in Hugging Face Diffusers from day one and handles text-to-image, multi-reference image-to-image, and bounding-box layout control in a single model. The distilled checkpoint generates a 1024 × 1024 image in about 2 seconds on an H100 using 16 steps with no CFG, while the base checkpoint uses 50 steps with CFG; NVFP4 and FP8 weights load via from_single_file for smaller GPUs, and LoRA training scripts ship for both text-to-image and image editing.

Image-to-Image • 6B • Updated • 3.24k • 39 https://huggingface.co/black-forest-labs/flux-3-image-klein Welcome FLUX.3 Image Klein in Diffusers: The latest open image generation model from Black Forect Labs Update on GitHub https://github.com/huggingface/blog/blob/main/flux-3.md Black Forest Labs https://huggingface.co/black-forest-labs FLUX.3 Image Klein is here, and it runs in Diffusers from day one. One model does text-to-image and multi-reference image-to-image, and it adds something new, the layout control . You write bounding boxes into the prompt to say where each part of the image should go, or which parts of a reference image to move. Both open checkpoints are in the FLUX.3 Image Klein collection https://huggingface.co/collections/black-forest-labs/flux-3-image-klein : | Checkpoint | What it is | Defaults | |---|---|---| | black-forest-labs/flux-3-image-klein https://huggingface.co/black-forest-labs/flux-3-image-klein | Time step distilled, for fast inference | 16 steps, no CFG | | black-forest-labs/flux-3-image-klein-base https://huggingface.co/black-forest-labs/flux-3-image-klein-base | Base model | 50 steps, CFG | You can head on to the Hugging Face space https://huggingface.co/spaces/black-forest-labs/flux-3-image-klein and generate some images right now TL;DR - One model for text-to-image, multi-reference image-to-image, and bounding-box layout control. - The distilled checkpoint makes a 1024 × 1024 image in about 2 seconds on an H100. - NVFP4 and FP8 weights load with from single file , for smaller GPUs. - LoRA training scripts for both text-to-image and image editing. - Ships as a Modular Diffusers modular-diffusers pipeline, so you can pop blocks out or add your own. Contents Getting started Install the latest diffusers and transformers for the Qwen3-VL text encoder : pip install -U diffusers transformers Load the distilled checkpoint and generate an image: python import torch from diffusers import ModularPipeline pipe = ModularPipeline.from pretrained "black-forest-labs/flux-3-image-klein" pipe.load components dtype=torch.bfloat16 pipe.to "cuda" pipe prompt="a red fox", output="images" 0 .save "fox.png" What FLUX.3 can do Text-to-image FLUX.3 responds best to a full scene description with the subject, the setting, the light, the composition, and the style. The official prompting guide https://docs.bfl.ai/guides/prompting summary from Black Forest Labs has more tips. The distilled checkpoint generates a 1024 × 1024 image in 16 steps by default. Pass height , width , or num inference steps to change them. prompt = "A wide-angle wildlife photograph of a red fox standing alert in a snow-covered forest clearing at golden hour. " "The fox is centered in the frame, its thick orange winter coat contrasting with the fresh white snow, breath " "visible in the cold air. Snow-laden pine trees fill the softly blurred background, and low warm sunlight rakes " "across the scene from the left, casting long blue shadows." fox = pipe prompt=prompt, generator=torch.Generator .manual seed 0 , output="images" 0 fox.save "fox.png" Image-to-image Pass a reference image or more with the image argument. Write the prompt as a caption of the image you want , with the change included. prompt = "The same red fox in the same snow-covered forest clearing, now at night under a full moon: cool blue " "moonlight, deep shadows, stars visible through the pine branches, the fox's eyes catching the light." night = pipe prompt=prompt, image=fox, from the previous text to image example generator=torch.Generator .manual seed 0 , output="images", 0 night.save "night.png" With several references, refer to them in the prompt as <ref image 0 , <ref image 1 , ... in the order they are passed to image : red fox = pipe prompt="a red fox", generator=torch.Generator .manual seed 0 , output="images" 0 bench = pipe prompt="a wooden bench", generator=torch.Generator .manual seed 0 , output="images" 0 combined = pipe prompt="The fox from <ref image 0 sitting on the wooden bench from <ref image 1 in a snowy setting.", image= red fox, bench , output="images", 0 combined.save "combined.png" Layout control Add a JSON list of regions after the caption. Each region has an id , a bbox , and a desc , and the caption mentions it as <id . Boxes are written as y min, x min, y max, x max , y first , on a 0–1000 scale. Because the scale is normalized, the same box works at any image size. Placing regions in a new image layout prompt = "Minimalist graphic illustration featuring a black silhouette of a person <silhouette 1 centered against a " "solid, vibrant chartreuse background <background 1 . The figure is captured in a dynamic, mid-stride running " "pose, facing toward the left side of the frame. " ' {"id": "background 1", "bbox": 0, 0, 1000, 1000 , "desc": "A flat field of neon yellow-green with a subtle ' 'paper texture and fine, organic grain."}, {"id": "silhouette 1", "bbox": 150, 150, 850, 850 , "desc": "A ' "black silhouette of a person in motion, made of dense stippled ink. The leading edges are solid; the trailing " 'arm and leg dissolve into coarse square pixels and scattered dots."} ' runner = pipe prompt=layout prompt, generator=torch.Generator .manual seed 0 , output="images" 0 Moving regions within a reference image For edits, each region says where its content comes from src bbox in the reference and where it goes tgt bbox in the output . A region with the same two boxes stays where it is. This example moves the knitted knight up and to the left, and pins the backdrop, the cliff, the needle, the dragon, and the fire: python from diffusers.utils import load image knight image = load image "knight orig.png" edit prompt = "In <ref image 0 , move the miniature grey amigurumi knight figure <knight 1 upwards and to the left along the " "yarn cliff <cliff 1 . Keep the rest of the image completely unchanged, preserving the blurred yarn backdrop " "<backdrop 1 , the large knitted dragon <dragon 1 , and the fiery orange thread <fire 1 . " ' {"id": "knight 1", "from": "ref image 0", "src bbox": 500, 150, 850, 350 , "tgt bbox": 194, 55, 544, 255 , ' '"desc": "A miniature amigurumi knight crocheted from thick grey woolen yarn."}, ' '{"id": "backdrop 1", "from": "ref image 0", "src bbox": 0, 0, 1000, 1000 , "tgt bbox": 0, 0, 1000, 1000 , ' '"desc": "A soft, out-of-focus backdrop of cream and indigo knitted yarn."}, ' '{"id": "cliff 1", "from": "ref image 0", "src bbox": 300, 0, 1000, 450 , "tgt bbox": 300, 0, 1000, 450 , ' '"desc": "A steep cliff of stacked yarn skeins in mustard, terracotta, and indigo."}, ' '{"id": "needle 1", "from": "ref image 0", "src bbox": 420, 250, 550, 450 , "tgt bbox": 420, 250, 550, 450 , ' '"desc": "An oversized polished steel tapestry needle."}, ' '{"id": "dragon 1", "from": "ref image 0", "src bbox": 100, 600, 950, 1000 , "tgt bbox": 100, 600, 950, 1000 , ' '"desc": "A large, fluffy knitted dragon in terracotta and mustard yarn."}, ' '{"id": "fire 1", "from": "ref image 0", "src bbox": 350, 400, 650, 650 , "tgt bbox": 350, 400, 650, 650 , ' '"desc": "Curled puffs of fiery orange angora thread suspended in mid-air."} ' moved = pipe prompt=edit prompt, image=knight image, generator=torch.Generator .manual seed 0 , output="images" 0 Base model and quantized variants Base model Everything above uses the distilled checkpoint. The base checkpoint runs 50 steps with classifier-free guidance, and it's the one to fine-tune see LoRA finetuning lora-finetuning . pipe = ModularPipeline.from pretrained "black-forest-labs/flux-3-image-klein-base" pipe.load components dtype={"vae": torch.float32, "default": torch.bfloat16} pipe.to "cuda" pipe.guider = pipe.guider.new guidance scale=5.0 optional image = pipe prompt=prompt, output="images" 0 Single-file format checkpoints Black Forest Labs also ships the transformer as a single file, in bf16 and in two quantized formats: fp8r rowwise FP8 and nvfp4 . Load any of them with from single file , then load the other components by name so the bf16 transformer isn't downloaded too. python from diffusers import BFLQuantizationConfig, Flux3Transformer2DModel, ModularPipeline model id = "black-forest-labs/flux-3-image-klein-base" variant = "fp8r" or "nvfp4" transformer = Flux3Transformer2DModel.from single file f"https://huggingface.co/{model id}/blob/main/variants/{variant}/model.safetensors", config=model id, subfolder="transformer", quantization config=BFLQuantizationConfig compute dtype=torch.bfloat16 , dtype=torch.bfloat16, pipe = ModularPipeline.from pretrained model id pipe.load components dtype=torch.bfloat16 pipe.update components transformer=transformer pipe.to "cuda" For the unquantized single file, use https://huggingface.co/{model id}/blob/main/model.safetensors and drop quantization config . | Variant | Fast kernels on | Other GPUs | |---|---|---| | fp8r | Compute capability 8.9+ Ada, Hopper, Blackwell | Weights are dequantized each step: lower memory, no speedup | | nvfp4 | Blackwell, with flashinfer-python and flashinfer-cubin at the same version | Same as above | LoRA finetuning FLUX.3 works for both text-to-image and image-to-image, which makes it a great base for custom LoRAs. diffusers ships DreamBooth LoRA training scripts for both tasks you can also train with Ostris' AI Toolkit https://github.com/ostris/ai-toolkit . Here is a text-to-image example that learns a tarot-card style from the multimodalart/1920-raider-waite-tarot-public-domain https://huggingface.co/datasets/multimodalart/1920-raider-waite-tarot-public-domain dataset, triggered with trtcrd : accelerate launch train dreambooth lora flux3 klein.py \ --pretrained model name or path="black-forest-labs/flux-3-image-klein-base" \ --dataset name="multimodalart/1920-raider-waite-tarot-public-domain" \ --caption column="caption" \ --instance prompt="trtcrd" \ --output dir="trained-flux3-tarot-lora" \ --mixed precision="bf16" \ --resolution=1024 \ --use aspect ratio buckets \ --train batch size=1 \ --rank=16 \ --lora alpha=16 \ --learning rate=1e-4 \ --lr scheduler="constant" \ --max train steps=1500 \ --validation prompt="a trtcrd of a cat, tarot style" \ --validation epochs=25 \ --seed=0 \ --push to hub To reduce memory usage, add a few optimization flags: --gradient checkpointing \ --cache latents \ --use 8bit adam --gradient checkpointing recomputes activations in the backward pass lower VRAM, slightly slower , --cache latents pre-encodes the images and frees the VAE, and --use 8bit adam keeps the optimizer states in 8-bit. A couple of knobs are worth experimenting with. The grid below trains the same tarot style with different LoRA ranks and optimizers; the adapters are on the Hub at diffusers-modular/flux3-tarot-lora https://huggingface.co/diffusers-modular/flux3-tarot-lora . More training features and configurations are documented in the DreamBooth FLUX.3 README https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README flux3 klein.md . For instruction-based editing, the image-to-image script train dreambooth lora flux3 klein img2img.py trains on reference image, target image, instruction triplets, where the reference is addressed in the prompt as <ref image 0 . See the DreamBooth FLUX.3 README https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README flux3 klein.md for all options and the full image-to-image example. What it costs to run In this section, we will talk about the memory consumption and the speed of generation of the models. Memory | name | peak allocated after load GiBs | peak allocated after inference GiBs | |---|---|---| | nvfp4 RTX6000 | 13.4 | 15.3 | | fp8r H100 | 14.35 | 16.27 | | bf16 H100 | 18.80 | 20.72 | If you have less memory: - Load a quantized transformer base-model-and-quantized-variants . - Run the text encoder separately and free it. It only runs once per image, before denoising see Encode once, generate many encode-once-generate-many . - Turn on automatic CPU offloading, which moves each component to the GPU only while it runs we didn't benchmark this python from diffusers import ComponentsManager, ModularPipeline manager = ComponentsManager manager.enable auto cpu offload device="cuda" pipe = ModularPipeline.from pretrained "black-forest-labs/flux-3-image-klein", components manager=manager pipe.load components dtype={"vae": torch.float32, "default": torch.bfloat16} no pipe.to "cuda" : the manager handles device placement Speed End-to-end timings for the distilled checkpoint 16 steps , averaged over three runs after a warmup: | Workflow | Output size | Time | Peak memory | |---|---|---|---| | Text-to-image | 1024 × 1024 | 1.93 s | 23.8 GiB | | Image-to-image, one 1024 × 1024 reference | 1024 × 1024 | 3.87 s | 23.8 GiB | | Layout editing the knight , one 1360 × 768 reference | 1360 × 768 | 4.10 s | 23.8 GiB | Single NVIDIA H100 80GB, no offloading, default attention backend. A reference image roughly doubles the time. Reference tokens join the same sequence as the text and the image being generated, so a 1024 × 1024 reference adds 4,096 tokens to every transformer call | Stacked bars of tokens per transformer call: text-to-image 4,256 tokens; image-to-image 8,272; layout text-to-image 4,336; layout image-to-image 8,800 | The architecture Let's open the model up, one piece at a time. The model takes three kinds of tokens: 1. text 2. the noisy image being generated 3. and optional reference images From pixels to tokens FLUX.3 is a latent-space flow model for image generation. This means that the denoiser takes latents as inputs and produces latents as outputs. We use a VAE to decode the latents back to pixels. FLUX.3 uses the same VAE architecture as Flux.2. The FLUX.2 VAE compresses an image 8× in each direction into 32 channels, and then every 2 × 2 patch of that latent is packed into a single token with 128 channels. So one token covers a 16 × 16 pixel patch , and a 1024 × 1024 image becomes a 64 × 64 grid of 4,096 tokens. The noisy image starts as random noise of exactly that shape. A reference image goes through the same VAE encoder and packing, and keeps its own size. From prompt to tokens The text encoder is Qwen3-VL, a vision-language model. FLUX.3 doesn't just take its last layer. It reads the hidden states of eight of its 36 layers layers 4, 8, …, 32 and stacks them, so each text token becomes a 8 × 2560 = 20,480-dimensional vector before a linear layer projects it to the transformer's width of 3072. This gives the transformer features from several depths of the language model, not just its final layer. The prompt is padded to a multiple of 80 tokens, and the transformer attends to the padding too, which is why the pipeline takes one prompt per call. Two stages The transformer processes the tokens in two stages: 1. Stage 1: per-mode blocks. Each stream goes through 5 blocks with its own weights. Text has its own blocks, the noisy image has its own, and the reference images have their own. 2. Stage 2: joint blocks. The streams are concatenated into one sequence, text | image | reference , which goes through 28 blocks that share one set of weights. The two stages differ in who can attend to whom . In stage 1, each stream only attends to itself, so each kind of token gets refined on its own terms first. In stage 2, everything attends to everything, where the image learns from the text and from the references. For comparison, FLUX.2's double-stream blocks also give text and image separate weights, but let them attend to each other from the start. FLUX.3 keeps the streams fully apart first, then merges them. At the end, the text and reference tokens are simply dropped. Only the image tokens go through the output layers, which predict how to move them one step toward a clean image. Inside one block Every block, in either stage, has the same shape. Attention and the MLP run in parallel on the same normalized input, and their sum is added back to the residual stream. Each block starts with a LayerNorm that has no learned scale or shift of its own. The scale, shift, and gate come from a timestep embedding instead this is called modulation . Inside the attention, the queries and keys are RMS-normalized and get rotary position embeddings RoPE , which we look at below. In a joint block, all tokens share the same weights, but text, image, and reference tokens each get their own scale, shift, and gate. python def block x, shift, scale, gate : shift, scale, gate come from the token's timestep embedding h = layer norm x no learned scale or shift h = 1 + scale h + shift q, k, v = to q h , to k h , to v h q, k = rope rms norm q , rope rms norm k out = attention q, k, v + mlp h parallel, MLP is SwiGLU return x + gate out Every token gets its own timestep In FLUX.1 and FLUX.2, one timestep embedding modulates every token. In FLUX.3, the modulation depends on the kind of token. The noisy image tokens are modulated by the current timestep t . The text and reference tokens are always modulated by timestep 0 , which means "clean, nothing to denoise here". So even a reference that looks almost the same as the output is marked as given context, not as something to generate. The distilled checkpoint is a flow map . It is also told the timestep each step jumps to. That second timestep gets its own embedding, which is added to every token's modulation, so each of its 16 steps can make a larger jump. Positions in four dimensions Attention on its own doesn't know where a token is this phenomenon is called permutation equivariance . RoPE fixes that by rotating each query and key by an amount that depends on the token's position. FLUX.3 uses four position axes, t, h, w, l timestep, height, width, length of sequence , and gives each one a quarter 32 dimensions of every 128-dimensional attention head: | Tokens | t: timestep | h: height | w: width | l: length of sequence | |---|---|---|---|---| | text token i | 0 | 0 | 0 | i | | image token at row, col | 0 | row / 2 | col / 2 | 0 | | reference k at row, col | 0 | row / 2 | col / 2 | − k + 1 | Text tokens line up along l , like words in a sentence. The image and the references are 2D grids in h, w , separated from each other along l . Because a reference uses the same h, w coordinates as the output, "the same spot in the reference" has a precise meaning to the model, which is what the src bbox / tgt bbox editing relies on. The denoising loop FLUX.3 is a flow-matching model. Generation starts from pure noise, and at each of the 50 steps the transformer predicts a velocity . This is the direction from noise toward a clean image. The scheduler takes one step along it. At every step you can also read off the model's current guess of the final image. Decoding that guess at a few steps shows how the image takes shape: The fox, its pose, and the light are already in place by step 10, while the noise level is still 0.97. The snow-laden branches and the texture of the fur only appear in the last few steps. FLUX.3's scheduler uses a large time shift 6.93 , which keeps the noise level high for most of the run: 44 of the 50 steps happen above σ = 0.5. We built this figure by adding one small block to the pipeline, which brings us to Modular Diffusers. If you are new to flow matching, we have a beginner friendly tutorial https://huggingface.co/spaces/ariG23498/flow-matching and a video https://youtu.be/hXQZ6lyRCsI for you to understand how it works. Modular Diffusers FLUX.3 ships as a Modular Diffusers https://huggingface.co/docs/diffusers/main/en/modular diffusers/overview pipeline. See the workflows You can list the workflows and print the exact blocks of each one, without loading any weights. python from diffusers import Flux3AutoBlocks blocks = Flux3AutoBlocks print blocks.available workflows 'text2image', 'image conditioned' print list blocks.get workflow "text2image" .sub blocks 'text encoder', 'denoise.input', 'denoise.prepare latents', 'denoise.set timesteps', 'denoise.prepare rope inputs', 'denoise.denoise', 'denoise.after denoise', 'decode' get workflow resolves the auto branches and gives you the plain sequence of steps that will run. print blocks.get workflow "image conditioned" .doc prints the components, inputs, defaults, and outputs of that workflow. Encode once, generate many Pop the text encoder out, encode the prompt once, free it, and generate from the cached embeddings. python import torch from diffusers import ModularPipeline model id = "black-forest-labs/flux-3-image-klein-base" blocks = ModularPipeline.from pretrained model id .blocks text blocks = blocks.sub blocks.pop "text encoder" blocks now starts at the VAE encoder text pipe = text blocks.init pipeline model id text pipe.load components dtype=torch.bfloat16 text pipe.to "cuda" state = text pipe prompt=prompt, negative prompt="" no guider here, so ask for the negative embeddings prompt embeds = state.get "prompt embeds" negative prompt embeds = state.get "negative prompt embeds" del text pipe torch.cuda.empty cache image pipe = blocks.init pipeline model id no text encoder to load image pipe.load components dtype={"vae": torch.float32, "default": torch.bfloat16} image pipe.to "cuda" for seed in range 3 : image = image pipe prompt embeds=prompt embeds, negative prompt embeds=negative prompt embeds, generator=torch.Generator .manual seed seed , output="images", 0 image.save f"{seed}.png" Add a step inside the loop The denoising figure above came from one extra block that records the model's guess of the clean image at every step: python from diffusers.modular pipelines import Flux3AutoBlocks, InputParam, ModularPipelineBlocks, OutputParam class Flux3PreviewStep ModularPipelineBlocks : model name = "flux3" @property def inputs self : return InputParam "latents" , InputParam "noise pred" @property def intermediate outputs self : return OutputParam "x0 previews", type hint=list @torch.no grad def call self, components, block state, i, t : if i == 0: block state.x0 previews = sigma = t / 1000 timesteps are sigmas scaled to 0, 1000 block state.x0 previews.append block state.latents - sigma block state.noise pred x0 = x t - sigma v return components, block state blocks = Flux3AutoBlocks denoise loop = blocks.sub blocks "denoise" .sub blocks "text2image" .sub blocks "denoise" denoise loop.sub blocks.insert "preview", Flux3PreviewStep , 1 between "denoiser" and "after denoiser" pipe = blocks.init pipeline model id pipe.load components dtype={"vae": torch.float32, "default": torch.bfloat16} pipe.to "cuda" out = pipe prompt=prompt, generator=torch.Generator .manual seed 0 , output= "x0 previews", "images" Try it - Models: FLUX.3 Image Klein collection https://huggingface.co/collections/black-forest-labs/flux-3-image-klein - Docs: FLUX.3 in Diffusers https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux3 - Fine-tune: DreamBooth LoRA scripts https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README flux3 klein.md - Prompting: Black Forest Labs prompting guide https://docs.bfl.ai/guides/prompting summary Built something with FLUX.3? Share your LoRAs and pipelines on the Hub. We'd love to see them 😎