Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, and it is the first FLUX 3 image model you can actually download and run on your own PC. It is a 5.6B parameter open weights model that does text-to-image, image editing and multi-reference generation with up to 10 reference images, and ComfyUI had core support merged the same day. The headline claim is that the INT8 build runs on a GeForce RTX card with as little as 8 GB of VRAM. This guide walks through exactly which files to download for FLUX 3 Image Klein in ComfyUI, which folder each one goes in, which of the four official templates to start with, the settings that matter, how much disk and VRAM you really need, and the problems you are most likely to hit on day one.
- FLUX 3 Image Klein is a 5.6B open weights image model from Black Forest Labs (released October 15, 2026) for text-to-image, editing and up to 10 reference images, with native 4K output and bounding-box layout control.
- For an 8 to 12 GB GPU, download the INT8 step-distilled transformer, the INT8 Qwen3-VL-4B text encoder and the FLUX.2 VAE from the gated
black-forest-labs/flux-3-image-klein-comfyrepo. - Update ComfyUI to a build newer than the last release tag: support landed in PR #17064 after the tag.
- Distilled model: 16 steps, no guidance. Base model: 50 steps, guidance 3.5, and it is meant for fine-tuning and LoRA training.
- The license is the FLUX Kommunity License: free commercial use for individuals and companies under $5M annual revenue, according to BFL’s launch post. Read LICENSE.md before you sell anything.
What Is FLUX 3 Image Klein?
FLUX 3 Image Klein is the smaller, open weights member of Black Forest Labs’ FLUX 3 family. BFL calls it the second open weights release in FLUX 3, after FLUX 3 Action (a 7B robot-control model), but it is the first one that matters for people making images. The full FLUX 3 Image model that BFL showed on October 8 stays behind the API. Klein is what you can run at home.
According to BFL’s announcement, Klein shares the core capabilities of the full FLUX 3 Image model:
- Text-to-image generation with native 4K output and multiple aspect ratios.
- Image editing, with what BFL calls pixel restoration, which keeps unedited regions unchanged.
- Multi-reference generation with up to 10 reference images. In the prompt you refer to them as “image 1”, “image 2” and so on.
- Bounding boxes that set where an element goes, or limit an edit to a defined area. The boxes use a 0 to 1000 coordinate grid and guide placement and scale. They are not hard pixel masks.
The point of Klein is size. At 5.6B parameters for the image generator, it is smaller than Qwen-Image 2.1 (7B) and Krea 2 (12.9B), and BFL says it has 38% fewer parameters than FLUX.2 [klein] 9B while scoring higher. Smaller means it fits on more consumer GPUs and is cheaper to fine-tune, which is exactly why the ComfyUI community jumped on it.
Base vs Distilled: Which Checkpoint Do You Need?
BFL ships two checkpoints, and picking the wrong one is the most common reason people say Klein is “slow” or “looks mushy”.
| Klein (step-distilled) | Klein Base | |
|---|---|---|
| Hugging Face repo | black-forest-labs/flux-3-image-klein |
black-forest-labs/flux-3-image-klein-base |
| Purpose | Fast, low-latency generation. Same checkpoint the BFL API serves. | Fine-tuning, LoRA training and research |
| Steps | 16 | 50 |
| Guidance | None (guidance-distilled) | 3.5 |
| ComfyUI file prefix | Files with _sd_ in the name |
Files without _sd_ |
| Best for | Everyday image making and editing | Training your own style or character LoRA |
If you just want pictures, use the distilled model. NVIDIA’s own measurements in BFL’s post put the base model at 50 steps anywhere from 1.7 seconds on a GB300 to 43.8 seconds on a DGX Spark for a 1024×1024 image, while the distilled model at 16 steps took 8.9 seconds on an RTX 4070. On a consumer card, that difference is the gap between “usable” and “go make a coffee”.
Hardware Requirements: VRAM, RAM and Disk
The official minimum is short: BFL says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. The model card lists an RTX 3090 as the documented supported floor, and it warns that “quantized packages need a loader that supports their format; their file size is not a measure of total GPU memory use.” In other words, 8 GB is possible with the right files and ComfyUI’s offloading, but it is not a promise for every resolution or every template.
Here is what the pieces weigh on disk, taken from the Hugging Face listings:
| Component | BF16 | Quantized |
|---|---|---|
| Image transformer (5.6B) | 11.1 GB | INT8 5.8 GB, FP8 about 5.6 GB, NVFP4 5.3 GB |
| Text encoder (Qwen3-VL-4B) | 8.9 GB | INT8 4.9 GB |
| VAE (FLUX.2 VAE) | 336 MB | No quantized version |
| Prompt upsampler (optional) | 18.8 GB | INT8 9.7 GB |
That gives three realistic setups:
- Smallest working pipeline (NVFP4 or INT8 transformer, INT8 encoder, VAE): roughly 10.5 to 11 GB on disk. This is the 8 to 12 GB VRAM route.
- Full BF16 pipeline: about 20.3 GB on disk. Comfortable on 24 GB cards.
- Everything including the BF16 upsampler: about 39 GB on disk. This is where the “39 GB” figure you may have seen on Reddit comes from. You do not need it.
Which quant for which card? NVFP4 is a Blackwell format, so it is the one to try first on RTX 50 series cards. INT8 is the safe default for RTX 30 and 40 series, and it is what the official ComfyUI templates load. FP8 sits in between and is mainly useful on 40 series and newer. For system RAM, plan on at least 32 GB if you are on an 8 GB card, because ComfyUI will park the text encoder in RAM while the transformer runs.
If you have been following my MiniMax H3 out of memory fixes, the same rule applies here: the problem is usually which model is still sitting in VRAM at the wrong moment, not the raw size of the transformer.
Files to Download and Where They Go
BFL published a ComfyUI-ready pack at black-forest-labs/flux-3-image-klein-comfy. All Klein repos are gated, so log in to Hugging Face and accept the terms on each model page before you try to download, or you will get a 401 or 403 error.
| ComfyUI folder | File | Size | Need it? |
|---|---|---|---|
models/diffusion_models |
flux3_image_klein_sd_int8_convrot (distilled, INT8) |
5.8 GB | Yes, for most people |
models/diffusion_models |
flux3_image_klein_sd_nvfp4 (distilled, NVFP4) |
5.3 GB | Optional, RTX 50 series |
models/diffusion_models |
flux3_image_klein_int8_convrot (base, INT8) |
5.8 GB | Only for the base templates or training |
models/diffusion_models |
flux3_image_klein_bf16 / flux3_image_klein_sd_bf16 |
11.1 GB | Only on 24 GB+ cards |
models/text_encoders |
qwen3vl_4b_int8_convrot |
4.9 GB | Yes |
models/text_encoders |
qwen3vl_4b_bf16 |
8.9 GB | Optional, higher precision |
models/text_encoders |
qwen3.5_9b_flux3_prompt_upsampler_int8_convrot |
9.7 GB | Optional, see below |
models/vae |
flux2-vae.safetensors |
336 MB | Yes (same VAE as FLUX.2) |
The file names above follow the BFL ComfyUI pack as listed by ComfyUI Wiki. Check the exact names and extensions in the repo when you download, because packs are sometimes re-uploaded with small naming changes in the first week after a launch.
Good news if you already run FLUX.2: Klein reuses the FLUX.2 VAE. If flux2-vae.safetensors is already in your models/vae folder, you can skip that download.
Install FLUX 3 Image Klein in ComfyUI Step by Step
Step 1: Update ComfyUI past the last release tag
Core support was merged on October 15 in Comfy-Org/ComfyUI PR #17064 (by Kijai). It adds the FLUX 3 image DiT, the Qwen3-VL-4B text encoder setup and a layout prompt node for bounding boxes. Because the merge happened after the latest release tag, a “stable” update may not include it yet. Use one of these:
- Portable Windows build: run
updateupdate_comfyui.bat(not the stable variant), then restart. - Git install:
git pullin the ComfyUI folder, thenpip install -r requirements.txtinside your venv. - ComfyUI Manager: Update ComfyUI, and make sure the channel is set to nightly or latest rather than stable.
Also update the workflow templates package (pip install -U comfyui-workflow-templates inside your venv), because the Klein templates ship in it.
Step 2: Accept the gated terms on Hugging Face
Open each repo you plan to download from (flux-3-image-klein-comfy, and the base or distilled repos if you download from those directly) and accept the license. If you download from the command line, log in first with hf auth login using a read token.
Step 3: Download the files
From the command line, a minimal 8 to 12 GB setup looks like this (adjust the include patterns to the exact names in the repo):
pip install -U "huggingface_hub[cli]"
hf download black-forest-labs/flux-3-image-klein-comfy
--include "*sd_int8_convrot*" "*qwen3vl_4b_int8_convrot*" "*flux2-vae*"
--local-dir ./klein-downloadThen move each file into the folder from the table above. If you prefer clicking, the “Files and versions” tab on Hugging Face works just as well.
Step 4: Load a template and run
Restart ComfyUI, open Workflow → Browse Templates, search for “Klein” and open FLUX 3 Image Klein Distilled Text to Image. If ComfyUI shows a missing models dialog, the file names in the loaders do not match what is on disk. Point each loader at your files and press Run.
The Four Official Workflow Templates
The comfyui-workflow-templates package includes four Klein graphs. All four reference the INT8 builds by default.
| Template | Checkpoint | Use it when |
|---|---|---|
| FLUX 3 Image Klein Distilled Text to Image | Distilled | Start here. Fast everyday generation. |
| FLUX 3 Image Klein Distilled Image Edit | Distilled | Edit a photo or render with a text instruction, optionally with reference images. |
| FLUX 3 Image Klein Text to Image | Base | Testing a LoRA trained on the base model, or comparing quality at 50 steps. |
| FLUX 3 Image Klein Image Edit | Base | Editing with a base-trained LoRA. |
BFL’s own ComfyUI pack also has a workflows/ folder. Those graphs are worth opening to see how BFL wires the bounding-box layout node and the optional upsampler.
Recommended Settings
| Setting | Distilled | Base | Notes |
|---|---|---|---|
| Steps | 16 | 50 | From BFL and the model card |
| Guidance | None | 3.5 | The distilled model is guidance-distilled; leave CFG at 1 or use the template’s value |
| Resolution to start | 1024×1024 (1 MP) | 1024×1024 | All official speed figures are at 1 MP. Move to 2 MP or 4K once 1 MP works. |
| Transformer | INT8 (NVFP4 on RTX 50) | INT8 | BF16 only with 24 GB+ |
| Text encoder | Qwen3-VL-4B INT8 | Qwen3-VL-4B INT8 | BF16 encoder if you have the VRAM |
| Upsampler | Off | Off | Turn on only for short, vague prompts |
For speed, BFL says the distilled model with INT8 generates a 1 MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 in ComfyUI. Those are vendor numbers measured by NVIDIA, and as of this writing nobody outside NVIDIA or BFL has published a full reproduction, so treat them as best-case. Adding a LoRA costs roughly 5% extra generation time, according to the same post.
Image Editing, Multi-Reference and Bounding Boxes
Editing
The Distilled Image Edit template takes an input image and an instruction. Klein’s pixel restoration is designed to keep areas you did not ask to change identical to the source, which is what makes it useful for product shots and thumbnails. Keep the instruction specific: “replace the mug with a glass of orange juice, keep the lighting” works better than “make it better”.
Multi-reference (up to 10 images)
Load up to ten reference images and refer to them in order inside the prompt, for example: “the woman from image 1 wearing the jacket from image 2, standing in the street from image 3, golden hour”. This is the feature that makes Klein interesting for consistent characters, the same problem I solved for video in the MiniMax H3 reference workflow. More references means more tokens and more memory, so on an 8 GB card start with two or three.
Bounding boxes
The layout prompt node added in PR #17064 lets you describe where things go with boxes on a 0 to 1000 grid, so [0, 0, 500, 1000] is the left half of the frame regardless of the final resolution. Boxes guide placement and scale. They are not masks, so do not expect edges to land on the exact pixel. BFL’s GitHub also has a Gradio demo for drawing boxes, which is handy for working out coordinates before you type them into ComfyUI.
Do You Need the Prompt Upsampler?
BFL ships a separate prompt upsampler, a Qwen3.5 9B model tuned for FLUX 3 (black-forest-labs/flux-3-image-klein-pup). It expands a short prompt into a long, detailed one before generation. The BFL API runs it for you, and the reference command line tool runs it on every request unless you pass --no-pup.
In ComfyUI it is optional, and for most local users I would skip it. It is 9.7 GB even in INT8, it adds another model to swap in and out of VRAM, and if you already write detailed prompts it has little to add. It does make sense if you like typing three-word prompts, or if you want your local results to look closer to what the API returns for the same prompt.
How It Compares: Benchmarks and Speed Claims
All of these are BFL’s internal evaluations, so read them as marketing until independent tests arrive:
- BFL says Klein ranks first among the models it tested in text-to-image, multi-reference generation and image editing, ahead of larger open weights models.
- Against FLUX.2 [klein] 9B it claims +91 Elo in text-to-image and +70 Elo in image editing, with 38% fewer parameters.
- It also claims higher ratings than the closed Ideogram V4.5 and FLUX.2 [pro].
- Krea 2 was left out of the editing comparison because it does not support editing.
- Median API latency for a 1 MP image: 4.54 seconds for text-to-image and 7.18 seconds for editing, including queueing on BFL’s servers.
The honest summary: Klein is clearly a step up from FLUX.2 [klein] on paper, it is the smallest of the current top open image models, and it is the only one of them with day-one native ComfyUI templates for both generation and editing. Whether it beats Qwen-Image 2.1 on your specific prompts is something to test yourself.
The FLUX Kommunity License in Plain English
Klein is released under the FLUX Kommunity License (BFL’s spelling). According to BFL’s launch post:
- Individuals and companies with annual revenue under $5M can use it commercially for free.
- Research and non-commercial use are permitted.
- Larger organizations need a self-hosted commercial license or the API.
One caution: write-ups have pointed out that the model card wording differs slightly from the launch post. If you plan to sell images, train and sell LoRAs, or build a product on Klein, read LICENSE.md in the repo itself rather than relying on a summary, including this one.
Troubleshooting Common Problems
| Symptom | Likely cause | Fix |
|---|---|---|
| Template not in the list, or red “missing node” boxes | ComfyUI or the templates package is older than PR #17064 | Update ComfyUI to nightly/latest and update comfyui-workflow-templates |
| 401 or 403 when downloading | Gated repo, terms not accepted | Accept the terms on each repo page, then hf auth login |
| Loader error on an INT8 or NVFP4 file | Loader does not support the quant format | Use the loader from the official template; NVFP4 needs a Blackwell (RTX 50) GPU |
| CUDA out of memory at 8 GB | Text encoder and transformer resident together, or resolution too high | Use the INT8 encoder, start at 1 MP, close GPU-heavy apps, launch with --reserve-vram 1.0 |
| Very slow, 40+ seconds per image | Using the base checkpoint at 50 steps | Switch to a _sd_ distilled file and the Distilled template (16 steps) |
| Washed out or over-cooked images on the distilled model | CFG or guidance set as if it were the base model | Reset guidance to the template default; the distilled model expects no guidance |
| Wrong colours or noise | Wrong VAE | Use flux2-vae.safetensors, not a FLUX.1 VAE |
| Bounding box items land off target | Coordinates entered in pixels | Use the 0 to 1000 grid, not pixel values |
Frequently Asked Questions
Can FLUX 3 Image Klein run on an 8 GB GPU?
BFL says the INT8 build runs on GeForce RTX GPUs with as little as 8 GB of VRAM via ComfyUI or Diffusers. Use the INT8 distilled transformer and INT8 text encoder, start at 1 MP, and have at least 32 GB of system RAM for offloading.
Is FLUX 3 Image Klein free for commercial use?
Under the FLUX Kommunity License, BFL’s launch post says commercial use is free for individuals and companies with under $5M in annual revenue. Larger companies need a separate license or the API. Check LICENSE.md for the exact wording.
Is there a GGUF version of FLUX 3 Image Klein?
Not officially. BFL ships BF16, FP8, INT8 and NVFP4 builds. Community GGUF conversions may appear, but the official INT8 build is already small and uses native kernels, which is usually faster than GGUF.
How many steps should I use?
16 steps for the distilled model with no guidance, and 50 steps with guidance 3.5 for the base model.
Does Klein use the FLUX.2 VAE?
Yes. Klein uses flux2-vae.safetensors (336 MB) in the models/vae folder, so existing FLUX.2 users already have it.
Can I train a LoRA on FLUX 3 Image Klein?
Yes. Train on the base checkpoint. BFL lists day-one LoRA support in Hugging Face Diffusers and Ostris’s AI Toolkit, plus NVIDIA NeMo AutoModel recipes.
Is Klein better than Qwen-Image 2.1?
BFL’s internal benchmarks rank Klein first, and it is smaller (5.6B vs 7B). Independent comparisons are still thin, so test both on your own prompts before switching your whole pipeline.
Final Thoughts
FLUX 3 Image Klein is the most practical open image model release of October 2026 for people with consumer GPUs: small enough for 8 to 16 GB cards in INT8, fast in its distilled form, and supported natively in ComfyUI from day one with templates for both generation and editing. Grab the INT8 distilled transformer, the INT8 Qwen3-VL-4B encoder and the FLUX.2 VAE, update ComfyUI past PR #17064, and start with the Distilled Text to Image template at 1 MP.
If you are building a local creative stack, these guides pair well with Klein:
- Best Local AI Video Models in 2026, to turn your Klein images into video.
- Best MiniMax H3 Image-to-Video Workflow for ComfyUI, which takes a still like the ones Klein makes as its first frame.
- Local Background Removal with U2NETP on CPU, for cutting out Klein product shots.
- ComfyUI Pixal3D Workflow: Image to 3D, to turn a Klein render into a 3D model.