Open weights · ComfyUI guide

FLUX 3 Klein in ComfyUI: Run Black Forest Labs’ Open Model on Your GPU

FLUX 3 Image Klein is the 5.6B open-weight model Black Forest Labs released on October 15, 2026: one model for text-to-image, editing and combining up to 10 reference images, with layout boxes and native 4K. Its INT8 build runs on an 8 GB RTX card. This guide covers the files, the four ComfyUI templates, VRAM, Base vs Distilled and LoRA training. You can try it in the box below before downloading anything.

5.6B open weights
Distilled: 16 steps
Up to 10 reference images
Layout boxes
Native 4K
INT8 on 8 GB VRAM

Try FLUX 3 Klein online before you set up ComfyUI

Photos to edit or combine (optional, up to 4)

from 5 credits

The box opens Artcraft’s image studio on FLUX.3 Klein (distilled), served by fal with BFL’s prompt expansion on. Add photos to edit or combine them, up to 4 here. Pick 1K, 2K or 4K there; the price in credits is on the button before you generate, and a failed image is refunded.

What FLUX 3 Image Klein is

FLUX 3 Image Klein (FLUX.3 Klein for short) is the small open member of Black Forest Labs’ FLUX 3 family, released on October 15, 2026. A 5.6B image transformer reads prompts through a Qwen3-VL 4B text encoder and decodes with the FLUX.2 VAE. The same weights generate from text, edit a photo, compose up to 10 reference images, and take bounding boxes that say where each element goes or which part of a photo to change. It renders natively up to 4K.

It comes as two checkpoints. Klein (distilled) runs 16 steps without CFG and is the one for making images. Klein Base runs 50 steps at CFG 3.5 and exists for fine-tuning, LoRAs and research. Both are meant to be used with pup, BFL’s 9B prompt upsampler (a Qwen3.5-9B fine-tune, Apache 2.0), which turns a short prompt into the long caption the model was trained on.

BFL’s own benchmarks put it ahead of FLUX.2 [klein] 9B by 91 Elo for text-to-image and 70 for editing with 38% fewer parameters, and above the other open models it tested. Those are BFL’s numbers, not an independent ranking. Its known weak spots are small text, which can garble, and boxes, which guide rather than mask: a tiny object can drop out and pixels outside a box can still change.

FLUX 3 Klein vs FLUX.2 Klein

Better on quality and features; stricter on licence. The second point is why FLUX.2 Klein 4B is still searched.

	FLUX.2 [klein] (January 2026)	FLUX 3 Image Klein (October 2026)
Sizes	4B and 9B, each distilled and base	One size, 5.6B, distilled and Base
Licence	4B Apache 2.0; 9B non-commercial	FLUX Kommunity: free commercial use under $5M yearly revenue
Steps (distilled)	4	16 (ComfyUI allows 4–16)
Prompt upsampler	None; write long prompts yourself	pup 9B, which BFL says to keep on
References	Multi-reference editing	Up to 10 reference images
Layout boxes	No	Yes, to place, move, resize or remove
Resolution	About 1–4 MP	Native up to 4K
VRAM	4B about 13 GB (less as GGUF)	INT8 from 8 GB, plus the upsampler

FLUX 3 Klein examples with prompts

Black Forest Labs’ own outputs from the model’s reference code, each with the short prompt that was typed; the upsampler wrote the long caption the model saw, as it does in the box above. Most cards show the distilled and Base checkpoints on the same prompt and seed.

Same prompt, same seed: distilled at 16 steps on the left, Base at 50 steps and CFG 3.5 on the right. Distilled is the checkpoint for finished images; Base is the undistilled one you train on.

FLUX 3 Klein ComfyUI: A man walking his dog in the afternoon sun — A man walking his dog in the afternoon sunFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
A man walking with his dog, afternoon sun

FLUX 3 Klein ComfyUI: A man walking his dog in the afternoon sun (Base) — A man walking his dog in the afternoon sunFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
A man walking with his dog, afternoon sun

FLUX 3 Klein ComfyUI: A brutalist concrete church lit by one slot of daylight — A brutalist concrete church lit by one slot of daylightFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
the interior of a brutalist concrete church, one slot of daylight falling on raw concrete, an empty wooden pew, dust in the light, 4x5 large format photograph

FLUX 3 Klein ComfyUI: A brutalist concrete church lit by one slot of daylight (Base) — A brutalist concrete church lit by one slot of daylightFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
the interior of a brutalist concrete church, one slot of daylight falling on raw concrete, an empty wooden pew, dust in the light, 4x5 large format photograph

FLUX 3 Klein ComfyUI: A fashion editorial in an oversized grey wool coat — A fashion editorial in an oversized grey wool coatFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
fashion editorial, a model in an oversized grey wool coat against a mustard seamless, one hard key light from the left, crisp shadow on the paper, hands in pockets, Hasselblad, no text

FLUX 3 Klein ComfyUI: A fashion editorial in an oversized grey wool coat (Base) — A fashion editorial in an oversized grey wool coatFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
fashion editorial, a model in an oversized grey wool coat against a mustard seamless, one hard key light from the left, crisp shadow on the paper, hands in pockets, Hasselblad, no text

FLUX 3 Klein ComfyUI: Rain over a wooden bridge as a woodblock print — Rain over a wooden bridge as a woodblock printFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a woodblock print of sudden rain over a wooden bridge, Hiroshige register, indigo and sumi, the rain as fine parallel lines, figures under umbrellas, no text

FLUX 3 Klein ComfyUI: Rain over a wooden bridge as a woodblock print (Base) — Rain over a wooden bridge as a woodblock printFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a woodblock print of sudden rain over a wooden bridge, Hiroshige register, indigo and sumi, the rain as fine parallel lines, figures under umbrellas, no text

FLUX 3 Klein ComfyUI: A knitted soft-sculpture diorama, macro — A knitted soft-sculpture diorama, macroFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
Macro photograph of a chunky knitted soft sculpture diorama, lit by a low, directional light source from the left that casts long, deep shadows across the textures. Filling the upper background is a dark indigo knitted backdrop. Across the lower third, a chaotic bed of thick looped yarn forms a churning sea. Rising from this base are three massive, conical icebergs constructed of stark white cable-knit yarn, positioned in a rhythmic frontal row: one on the left <iceberg_1>, a slightly taller one on the right <iceberg_2>, and a bulky central peak <iceberg_3>. Clinging to the slopes of the outer icebergs is a rhythmic repetition of tiny, identical pale blue knitted houses. Suspended in the upper frame between the peaks is a slightly curling strip of cream-colored stockinette fabric, bearing thick embroidered text <En_Text_1>. Disrupting the icy white and blue palette, a giant, vibrant magenta knitted whale tail <whale_tail_1> bursts vertically from the center of the teal sea, its thick ribbed surface covered in a halo of stray wool fibers illuminated by the raking light.
Layout boxes (reference code format): [ {"id": "iceberg_1", "desc": "A conical soft sculpture made of stark white, chunky roving yarn in a thick cable-knit pattern. The left side is brightly lit, while the right side falls into deep, fuzzy shadow.", "bbox": [0.3, 0.05, 0.75, 0.35]}, {"id": "iceberg_2", "desc": "A tall, asymmetrical peak of white cable-knit yarn, featuring a few visibly dropped stitches near the base and slightly uneven tension in the cables.", "bbox": [0.25, 0.65, 0.75, 0.95]}, {"id": "iceberg_3", "desc": "The largest central iceberg, a bulky mound of white and pale grey ribbed knitting, dominating the midground with thick, vertical columns of stitches.", "bbox": [0.2, 0.3, 0.65, 0.7]}, {"id": "En_Text_1", "desc": "The exact words \"KNITTED ICEBERG VILLAGE\" embroidered in thick, raised navy-blue yarn across the cream banner. The lettering is slightly uneven and puckered, conforming to the physical stretch of the knitted fabric beneath it.", "bbox": [0.13, 0.22, 0.22, 0.78]}, {"id": "whale_tail_1", "desc": "A giant whale tail knitted from bright magenta and coral-pink yarn, bursting vertically from the teal sea. It features thick ribs, a slightly asymmetrical fluke, and a fuzzy halo of stray wool fibers caught in the directional light.", "bbox": [0.5, 0.38, 0.85, 0.62]} ]

FLUX 3 Klein ComfyUI: A knitted soft-sculpture diorama, macro (Base) — A knitted soft-sculpture diorama, macroFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
Macro photograph of a chunky knitted soft sculpture diorama, lit by a low, directional light source from the left that casts long, deep shadows across the textures. Filling the upper background is a dark indigo knitted backdrop. Across the lower third, a chaotic bed of thick looped yarn forms a churning sea. Rising from this base are three massive, conical icebergs constructed of stark white cable-knit yarn, positioned in a rhythmic frontal row: one on the left <iceberg_1>, a slightly taller one on the right <iceberg_2>, and a bulky central peak <iceberg_3>. Clinging to the slopes of the outer icebergs is a rhythmic repetition of tiny, identical pale blue knitted houses. Suspended in the upper frame between the peaks is a slightly curling strip of cream-colored stockinette fabric, bearing thick embroidered text <En_Text_1>. Disrupting the icy white and blue palette, a giant, vibrant magenta knitted whale tail <whale_tail_1> bursts vertically from the center of the teal sea, its thick ribbed surface covered in a halo of stray wool fibers illuminated by the raking light.
Layout boxes (reference code format): [ {"id": "iceberg_1", "desc": "A conical soft sculpture made of stark white, chunky roving yarn in a thick cable-knit pattern. The left side is brightly lit, while the right side falls into deep, fuzzy shadow.", "bbox": [0.3, 0.05, 0.75, 0.35]}, {"id": "iceberg_2", "desc": "A tall, asymmetrical peak of white cable-knit yarn, featuring a few visibly dropped stitches near the base and slightly uneven tension in the cables.", "bbox": [0.25, 0.65, 0.75, 0.95]}, {"id": "iceberg_3", "desc": "The largest central iceberg, a bulky mound of white and pale grey ribbed knitting, dominating the midground with thick, vertical columns of stitches.", "bbox": [0.2, 0.3, 0.65, 0.7]}, {"id": "En_Text_1", "desc": "The exact words \"KNITTED ICEBERG VILLAGE\" embroidered in thick, raised navy-blue yarn across the cream banner. The lettering is slightly uneven and puckered, conforming to the physical stretch of the knitted fabric beneath it.", "bbox": [0.13, 0.22, 0.22, 0.78]}, {"id": "whale_tail_1", "desc": "A giant whale tail knitted from bright magenta and coral-pink yarn, bursting vertically from the teal sea. It features thick ribs, a slightly asymmetrical fluke, and a fuzzy halo of stray wool fibers caught in the directional light.", "bbox": [0.5, 0.38, 0.85, 0.62]} ]

FLUX 3 Klein ComfyUI: A two-ink risograph poster of a lido swimmer — A two-ink risograph poster of a lido swimmerFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a risograph poster in two inks, fluorescent orange and blue, slightly misregistered, a swimmer mid-dive in flat shapes, the word 'LIDO' large at the bottom, no other text

FLUX 3 Klein ComfyUI: A two-ink risograph poster of a lido swimmer (Base) — A two-ink risograph poster of a lido swimmerFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a risograph poster in two inks, fluorescent orange and blue, slightly misregistered, a swimmer mid-dive in flat shapes, the word 'LIDO' large at the bottom, no other text

FLUX 3 Klein ComfyUI: A still life lit like a Dutch painting — A still life lit like a Dutch paintingFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a still life photographed in the register of a Dutch painting: a torn loaf, a pewter jug, three lemons with one peeled in a spiral, a knife, on a dark oak table, one window light from the left, black background, 4x5 large format

FLUX 3 Klein ComfyUI: A still life lit like a Dutch painting (Base) — A still life lit like a Dutch paintingFLUX.3 Klein distilled (16 steps) vs Base (50 steps), same seed · Source
Show prompt
a still life photographed in the register of a Dutch painting: a torn loaf, a pewter jug, three lemons with one peeled in a spiral, a knife, on a dark oak table, one window light from the left, black background, 4x5 large format

Boxes tie a description to a region, here in the reference code’s [x_min, y_min, x_max, y_max] format. The box above does not send boxes yet; the prompt alone still works there.

FLUX 3 Klein ComfyUI: Hot-air balloons, each placed by a layout box — Hot-air balloons, each placed by a layout boxFLUX.3 Klein distilled vs Base, layout boxes, same seed · Source
Show prompt
hot-air balloons over a green valley at sunrise
Layout boxes (reference code format): [ {"id": "obj_1", "desc": "a red and yellow striped hot-air balloon", "bbox": [0.08, 0.06, 0.48, 0.38]}, {"id": "obj_2", "desc": "a small blue hot-air balloon", "bbox": [0.2, 0.62, 0.42, 0.8]} ]

FLUX 3 Klein ComfyUI: Hot-air balloons, each placed by a layout box (Base) — Hot-air balloons, each placed by a layout boxFLUX.3 Klein distilled vs Base, layout boxes, same seed · Source
Show prompt
hot-air balloons over a green valley at sunrise
Layout boxes (reference code format): [ {"id": "obj_1", "desc": "a red and yellow striped hot-air balloon", "bbox": [0.08, 0.06, 0.48, 0.38]}, {"id": "obj_2", "desc": "a small blue hot-air balloon", "bbox": [0.2, 0.62, 0.42, 0.8]} ]

FLUX 3 Klein ComfyUI: A cactus added inside a marked box (input 1) — A cactus added inside a marked boxFLUX.3 Klein distilled vs Base, one target box, same seed · Source
Show prompt
add a small potted cactus in the marked spot
Layout boxes (reference code format): [ {"id": "obj_1", "desc": "a small potted cactus", "from": "target", "src_bbox": null, "tgt_bbox": [0.55, 0.62, 0.85, 0.85]} ]

FLUX 3 Klein ComfyUI: A cactus added inside a marked box — A cactus added inside a marked boxFLUX.3 Klein distilled vs Base, one target box, same seed · Source
Show prompt
add a small potted cactus in the marked spot
Layout boxes (reference code format): [ {"id": "obj_1", "desc": "a small potted cactus", "from": "target", "src_bbox": null, "tgt_bbox": [0.55, 0.62, 0.85, 0.85]} ]

ComfyUI template outputs and LoRA results

What the official ComfyUI templates produce, and what LoRAs trained on Klein Base change: each LoRA card shows the same seed with and without the LoRA.

The sample outputs of ComfyUI’s FLUX 3 Klein templates, all on the INT8 files. The two edits use a box, written as <obj_1> in the prompt.

FLUX 3 Klein ComfyUI: A retro hotel entrance at night — A retro hotel entrance at nightComfyUI template: FLUX.3 Klein Distilled text to image, INT8 · Source
Show prompt
A cinematic, Dutch angle night shot of a retro hotel entrance illuminated by neon cyan and warm orange lights. A bellhop in uniform pulls a gold luggage cart with suitcases across a wet sidewalk. Inside the brightly lit, teal-toned glass lobby, two women in vintage uniforms stand near a wooden desk alongside tropical plants and a seated figure. The scene features a moody, stylized teal-and-orange color palette with strong neon reflections. Absolutely no text, no typography, no letters, no words, no numbers, no logos, no watermark, no captions, pure imagery only.

FLUX 3 Klein ComfyUI: An athlete pressing his forehead to a ball — An athlete pressing his forehead to a ballComfyUI template: FLUX.3 Klein Base text to image, INT8 · Source
Show prompt
close up profile of a young male athlete with closed eyes pressing his sweaty forehead against a soccer ball, right hand gripping the ball, stark white background, athletic jacket, white ball with jagged red graphic patterns, high contrast studio lighting, highly detailed skin texture, sharp focus Absolutely no text, no typography, no letters, no words, no numbers, no logos, no watermark, no captions, pure imagery only.

FLUX 3 Klein ComfyUI: A silk gown recoloured inside one box (input 1) — A silk gown recoloured inside one boxComfyUI template: FLUX.3 Klein Base image edit, INT8, one box · Source
Show prompt
In <ref_image_0>, change the dress color <obj_1> to red
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A silk gown recoloured inside one box — A silk gown recoloured inside one boxComfyUI template: FLUX.3 Klein Base image edit, INT8, one box · Source
Show prompt
In <ref_image_0>, change the dress color <obj_1> to red
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: An earring swapped inside one box (input 1) — An earring swapped inside one boxComfyUI template: FLUX.3 Klein Distilled image edit, INT8, one box · Source
Show prompt
In <ref_image_0>, replace the earring <obj_1> with a single large openwork metal statement earring, one bold curved pendant, intricate hollow metal cutouts, layered metal wires, polished warm bronze finish, structured artistic shape, keep it attached to the correct ear position. Match the character's existing lighting, shadows and skin tone. No pearls, no pearl beads, no scattered small beads. Keep all other parts of the image completely unchanged.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: An earring swapped inside one box — An earring swapped inside one boxComfyUI template: FLUX.3 Klein Distilled image edit, INT8, one box · Source
Show prompt
In <ref_image_0>, replace the earring <obj_1> with a single large openwork metal statement earring, one bold curved pendant, intricate hollow metal cutouts, layered metal wires, polished warm bronze finish, structured artistic shape, keep it attached to the correct ear position. Match the character's existing lighting, shadows and skin tone. No pearls, no pearl beads, no scattered small beads. Keep all other parts of the image completely unchanged.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

Style LoRAs BFL trained on Klein Base for its docs; the phrase at the end (“shtm style”, “bilb style”, “ohwj style”) is the trigger. Left with the LoRA, right without.

FLUX 3 Klein ComfyUI: An astronomer at a brass telescope, trained style LoRA — An astronomer at a brass telescope, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
An old astronomer in a tower room looks through a large brass telescope at the night sky. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: An astronomer at a brass telescope, trained style LoRA (No LoRA) — An astronomer at a brass telescope, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
An old astronomer in a tower room looks through a large brass telescope at the night sky. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A knight riding toward a castle, trained style LoRA — A knight riding toward a castle, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A knight on a white horse rides through a dark pine forest toward a distant castle on a hill. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A knight riding toward a castle, trained style LoRA (No LoRA) — A knight riding toward a castle, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A knight on a white horse rides through a dark pine forest toward a distant castle on a hill. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A fox and a crow under a full moon, trained style LoRA — A fox and a crow under a full moon, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A fox and a crow sit on a fallen tree, talking, while a full moon rises behind them. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A fox and a crow under a full moon, trained style LoRA (No LoRA) — A fox and a crow under a full moon, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A fox and a crow sit on a fallen tree, talking, while a full moon rises behind them. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A sailing ship in a storm, trained style LoRA — A sailing ship in a storm, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A sailing ship caught in a storm with huge waves and lightning over the sea. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A sailing ship in a storm, trained style LoRA (No LoRA) — A sailing ship in a storm, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A sailing ship caught in a storm with huge waves and lightning over the sea. shtm style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: The same knight, a second style LoRA — The same knight, a second style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A knight on a white horse rides through a dark pine forest toward a distant castle on a hill. bilb style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: The same knight, a second style LoRA (No LoRA) — The same knight, a second style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
A knight on a white horse rides through a dark pine forest toward a distant castle on a hill. bilb style
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A striped lighthouse under a starry sky, trained style LoRA — A striped lighthouse under a starry sky, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
ohwj style, a lighthouse on a cliff, vibrant medium. A tall striped lighthouse stands on a rocky cliff above swirling turquoise waves, its beam spreading into ribbons of color across a starry sky covered in dots and spirals. Seagulls with patterned wings circle the tower.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: A striped lighthouse under a starry sky, trained style LoRA (No LoRA) — A striped lighthouse under a starry sky, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
ohwj style, a lighthouse on a cliff, vibrant medium. A tall striped lighthouse stands on a rocky cliff above swirling turquoise waves, its beam spreading into ribbons of color across a starry sky covered in dots and spirals. Seagulls with patterned wings circle the tower.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: An orange cat with a cup of tea, trained style LoRA — An orange cat with a cup of tea, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
ohwj style, a cat and a cup of tea, whimsical medium. A large orange cat with a coat of dots and swirls curls around a giant teacup on a patterned tablecloth, steam rising in spirals that turn into tiny birds.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein ComfyUI: An orange cat with a cup of tea, trained style LoRA (No LoRA) — An orange cat with a cup of tea, trained style LoRAFLUX.3 Klein Base + a LoRA trained by BFL · Source
Show prompt
ohwj style, a cat and a cup of tea, whimsical medium. A large orange cat with a coat of dots and swirls curls around a giant teacup on a patterned tablecloth, steam rising in spirals that turn into tiny birds.
This prompt needs something the box above does not send, such as a LoRA or layout boxes. Copy it into ComfyUI instead.

FLUX 3 Klein download: files for ComfyUI

The ComfyUI builds live in Black Forest Labs’ gated Hugging Face repo black-forest-labs/flux-3-image-klein-comfy. Accept the licence there, then download.

File	ComfyUI folder	What it is
flux3-image-klein-int8-convrot.safetensors	models/diffusion_models/	Klein distilled, INT8. For images and edits.
flux3-image-klein-base-int8-convrot.safetensors	models/diffusion_models/	Klein Base, INT8. For the Base templates and LoRA testing.
qwen3vl_4b_int8_convrot.safetensors	models/text_encoders/	Qwen3-VL 4B text encoder.
qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors	models/text_encoders/	pup, the 9B prompt upsampler, 4-bit weights.
flux2-vae.safetensors	models/vae/	The FLUX.2 VAE, shared with FLUX.2 models.

With more than 24 GB of VRAM you can swap in the BF16 builds (flux3-image-klein-bf16 and flux3-image-klein-base-bf16, about 11.2 GB each), which ComfyUI says follow long prompts more closely. FP8 (5.6 GB) and NVFP4 builds are on BFL’s main repo.

Where the FLUX 3 Klein files go

Place the files as below, then restart ComfyUI or press R to refresh the model lists.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── flux3-image-klein-int8-convrot.safetensors
    │   └── flux3-image-klein-base-int8-convrot.safetensors
    ├── text_encoders/
    │   ├── qwen3vl_4b_int8_convrot.safetensors
    │   └── qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors
    └── vae/
        └── flux2-vae.safetensors

The 4 official FLUX 3 Klein ComfyUI workflows

All four ship with ComfyUI 0.41.0 and later (search “FLUX 3 Image Klein” in Browse Templates). The defaults below are ComfyUI’s.

Template	Checkpoint	Steps	CFG	Upsampler
FLUX 3 Image Klein Distilled: Text to Image	Distilled	4–16 (16 default)	1, fixed	On by default
FLUX 3 Image Klein: Text to Image	Base	50	3.5	On by default
FLUX 3 Image Klein Distilled: Image Edit	Distilled	4–16	1	Write the full result yourself
FLUX 3 Image Klein: Image Edit	Base	50	1	Write the full result yourself

The edit templates resize references to about 1 megapixel. Keeping them at native size (up to about 16 MP) keeps more detail but costs more compute.

Run FLUX 3 Klein in ComfyUI, step by step

1
Update to ComfyUI 0.41.0 or later
The FLUX 3 nodes and templates arrived in 0.41.0 (the templates in the 0.41.1 backport). Older builds show missing nodes.
2
Accept the licence and download
Open black-forest-labs/flux-3-image-klein-comfy on Hugging Face, accept the FLUX Kommunity License, and download the five files above into their folders.
3
Start with the Distilled text-to-image template
It is the fastest and the one BFL tuned for everyday images. Leave enable_prompt_upsampler on, type a short prompt and queue it.
4
Try fewer steps once it works
The distilled model takes 4 to 16 steps in ComfyUI. 8 is a good draft setting; go back to 16 for finals.
5
Edit with the Image Edit template
Load your photo, then describe the finished picture rather than the change: “the same kitchen, the mug now full of coffee with a croissant beside it”. Mark a region with a box and write its label, such as <obj_1>, in the prompt.

FLUX 3 Klein VRAM: what runs on 8 GB, 12 GB and 24 GB

The generator is small; the 9B upsampler is the heavy part. ComfyUI loads them one after the other, which is how the INT8 set fits an 8 GB card.

Setup	VRAM	Example cards	Notes
INT8 set (ComfyUI templates)	8 GB and up	RTX 3070, 4060 Ti 8 GB	BFL’s headline figure. 12 GB leaves room for 2K and several references.
FP8 generator	12–16 GB	RTX 4070, 4080	5.6 GB of weights; native FP8 on RTX 40 and newer.
BF16 generator	24 GB and up	RTX 4090, 5090	11.2 GB of weights. ComfyUI recommends it above 24 GB.
NVFP4	Lowest	RTX 50 series	Blackwell-only 4-bit format.
Upsampler off	Saves the 9B model	Any of the above	Then write long, detailed prompts yourself; short ones get noticeably worse.

BFL’s timings at 1024², 16 steps: about 8.9 s on an RTX 4070, and about 4 s for INT8 on an RTX 5090. Its reference code without offloading peaks far higher (25–31 GB on a B200), so use ComfyUI or Diffusers offloading on consumer cards.

FLUX 3 Klein LoRA training

Same rule as other distilled models: train on Base, run on the distilled checkpoint.

1
Prepare 20–40 captioned images
One style or subject per LoRA, with a rare trigger phrase in each caption. BFL’s doc LoRAs used phrases like “shtm style”.
2
Pick a trainer
Ostris AI Toolkit, the Diffusers DreamBooth LoRA script and NVIDIA NeMo all support Klein Base. Expect to need a 24 GB card for comfortable training.
3
Train on Klein Base
Base is undistilled, so it learns cleanly. LoRAs trained on the distilled model tend to fight its 16-step distillation.
4
Load it on the distilled model
Add a LoRA loader to the Distilled template at around 0.75 strength, BFL’s suggested starting point, and put the trigger phrase in the prompt.

FLUX 3 Klein license: free commercial use under $5M revenue

The weights are released under the FLUX Kommunity License. Individuals, and companies with less than US$5 million in annual revenue, may use them commercially for free. A company above that threshold needs a self-hosting licence from Black Forest Labs, or uses the model through a paid API instead.

Two more conditions apply to everyone: outputs may not be used to train a competing model, and the weights are gated, so you accept the licence on Hugging Face before downloading. This is a change from FLUX.2 [klein] 4B, which was Apache 2.0 with no revenue limit. This summary is not legal advice; read the licence before building a product on it.

FLUX 3 Klein ComfyUI FAQ

Can FLUX 3 Klein run on 8 GB of VRAM?

Yes, with the INT8 files in ComfyUI’s templates, which load the generator and the prompt upsampler one after the other. 12 GB is more comfortable for 2K images and several reference photos. Below 8 GB, use the box on this page.

Which FLUX 3 Klein files do I download?

For ComfyUI: the distilled INT8 model (and the Base INT8 if you want the Base templates), the Qwen3-VL 4B text encoder, the w4a8 prompt upsampler and the FLUX.2 VAE, all from black-forest-labs/flux-3-image-klein-comfy on Hugging Face. BF16 builds are for cards with more than 24 GB.

Is FLUX 3 Klein free for commercial use?

Only below a revenue limit. Under the FLUX Kommunity License, individuals and companies with less than US$5 million in annual revenue can use it commercially for free; larger companies need a licence from Black Forest Labs or must use a paid API. Outputs may not be used to train competing models.

Do I need the prompt upsampler?

BFL says to keep it on: Klein was trained on long captions, and short prompts fed in directly come out noticeably worse. You can switch enable_prompt_upsampler off in ComfyUI to save memory or to control the wording, but then write long, detailed prompts.

Distilled or Base: which FLUX 3 Klein should I use?

Distilled for generating and editing: 16 steps, no CFG, about three times faster. Base for training LoRAs and fine-tunes: 50 steps at CFG 3.5, where negative guidance works and LoRAs learn cleanly.

How many reference images can FLUX 3 Klein take?

Up to 10 per edit, referred to in the prompt by order (“the first image”) or as <ref_image_0>, <ref_image_1> in ComfyUI. The box on this page sends up to 4.

How do FLUX 3 Klein bounding boxes work?

Each box pairs a region with a description; name it with a label in the prompt, such as <obj_1>. On text-to-image a box places a new element; on an edit it can add, move, resize or remove one. Boxes guide rather than mask, so small objects can be missed.

Is FLUX 3 Klein better than FLUX.2 Klein?

On quality and features, by BFL’s benchmarks: +91 Elo for text-to-image and +70 for editing against FLUX.2 [klein] 9B, plus 10 references, boxes and 4K. FLUX.2 Klein 4B still has the looser Apache 2.0 licence and a 4-step distilled model.

Which ComfyUI version supports FLUX 3 Klein?

ComfyUI 0.41.0 or later; the four templates ship from the 0.41.1 backport. Update before loading a FLUX 3 workflow, or the nodes show as missing.

Is Artcraft affiliated with Black Forest Labs?

No. Artcraft is an independent studio that runs FLUX.3 Klein through fal’s API. FLUX is a trademark of Black Forest Labs; the sample images on this page are BFL’s and ComfyUI’s, with links to where they were published.

Try FLUX 3 Klein without a GPU

Generate, edit and combine photos in the browser, up to 4K.

Try FLUX 3 Klein

Other Artcraft pages that open the same studio, set up for a different job.