DEV Community

AI-KSK
AI-KSK

Posted on Originally published at bilibili.com

FLUX 3 Klein Tested: 5.6B Open Weights, Native 4K, and Seven ComfyUI Workflows

#ai #opensource #tutorial #comfyui

A hands-on look at Black Forest Labs' newest open image model — where it wins, where it loses, and which of the seven workflows you should actually start with.

The short version

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, and it is genuinely open weights under the FLUX Kommunity License. That means free commercial use for individuals and companies under $5M annual revenue — not an API key, not a waitlist.

On paper it reads like a spec sheet:

  • 5.6B parameters — the smallest model in its comparison class
  • Native 4K, up to ~16 megapixels
  • Text-to-image, image editing, and multi-reference generation (up to 10 reference images)
  • Bounding boxes for layout control
  • 11.2 GB in BF16, 5.6 GB in FP8 — and INT8 builds run on 8 GB of VRAM

In a 100-image composite benchmark run for this test, FLUX 3 Klein scored 84.68 against Qwen Image 2.1's 85.72. A 1.04-point gap. That is not "catching up" territory anymore; it is a coin flip depending on the task.

My conclusion after running 28 comparison tests and seven workflows: use the distilled model as your default. The quality difference against the base model is smaller than the speed difference, and in a few task types the distilled path actually wins.

Base or distilled? The 28-test answer

BFL shipped two checkpoints:

Checkpoint Steps CFG Best for
flux3_image_klein_base 50 3.5 Fine-tuning, maximum fidelity, A/B reference
flux3_image_klein_sd (step-distilled) 16 1.0 Everything else

The distilled model is trained to run without classifier-free guidance — CFG stays at 1.0. That alone cuts sampling cost roughly in half before you even count the step reduction.

Where the base model still earns its keep:

  • Portraits — the base checkpoint renders slightly heavier facial texture, which reads as more photographic
  • Final deliveries where you can afford 50 steps and want maximum literal fidelity

Where the distilled model wins or matches:

  • Product text — subjectively cleaner rendering, fewer artifacts on typography
  • Pose editing — more stable across seeds
  • Iteration speed — 16 steps means you can explore 3x more seeds in the same wall time

Practical rule: start every task on the distilled checkpoint. Reach for base only when a specific output needs more literal detail, and use workflow 7 (below) to compare on identical seeds.

The seven workflows, and when to use each

The package covers the four official graphs plus three derived ones. They are numbered in a suggested learning order.

1. Text-to-Image (official, base model)

The canonical graph: text encode → latent → sample → decode → save.

  • Sampler: 50 steps, CFG 3.5, euler / simple
  • Resolution: set on EmptyFlux2LatentImage, or through a ResolutionSelector node
  • Includes CreateBoundingBoxes and Flux3LayoutPrompt stubs, so you can switch from plain prompting to layout-guided prompting without rebuilding the graph

Use this when you need the reference-quality output, or when you are comparing against the distilled graph.

2. Image Editing (official, base model)

The other official graph, and the one that shows what Klein can actually do.

  • Sampler: 50 steps, CFG 1
  • Bounding-box editing via CreateBoundingBoxes + Flux3LayoutPrompt
  • Multi-image input through ReferenceMultiLatent and BatchImagesNode — up to 10 references, addressed in the prompt as ref_image_0 through ref_image_9
  • ComfySwitchNode toggles edit variants without duplicating the graph
  • Prompt upsampler included (qwen3.5_9b_flux3_prompt_upsampler_w4a8) — bypass it when you already have a polished caption

The prompt rule that matters: describe the final frame, not the change. "Swap the earring" degrades badly. "The woman from ref_image_0, now wearing a silver hoop earring, same pose, same lighting" works.

3. Distilled Text-to-Image

The 9-node fast path. Same latent space and VAE as the base graph, so seeds and resolutions are interchangeable between the two — which is what makes fair comparison possible.

  • Sampler: 16 steps, no CFG guidance
  • Prompt upsampler optional
  • Nothing to configure beyond prompt and resolution

If you only learn one workflow, learn this one.

4. Distilled Image Editing

Same editing capability as graph 2, on the distilled checkpoint. This is the iteration workhorse: run five seeds here, pick the best, then optionally re-run it on the base graph for the final.

  • Sampler: 16 steps, no CFG guidance
  • Full bounding-box + multi-reference support
  • PrimitiveStringMultiline keeps long captions manageable

5. Multi-Reference 1 → 10

Feed one to ten reference images and assign each a job in the prompt.

  • Sampler: 16 steps, no CFG guidance
  • ReferenceMultiLatent packs references into a single conditioning tensor
  • ConditioningZeroOut lets you disable a slot without deleting its image — fast A/B of reference sets

The honest limitation: 1–2 references is the sweet spot. Accuracy degrades noticeably past 4 references. If you are stacking ten images, expect to fight for consistency.

6. Bounding-Box Layout Edit

The workflow that makes boxes do real work.

  • Sampler: 16 steps, no CFG guidance
  • CreateBoundingBoxes defines src_bbox / tgt_bbox pairs on a 0–1000 normalised grid, so the same box works at any output resolution
  • Flux3LayoutPrompt converts boxes into the caption syntax the model expects

The technique worth knowing: pin everything you do not want to move by giving it an identical source and target box. An unmoved box is a "keep this exactly here" instruction, and it is far more reliable than asking the model politely in prose.

One expectation to set: bounding boxes bias placement strongly but they are not a hard mask. Expect small drift, especially on organic shapes.

7. Base vs Distilled A/B

Two checkpoints in one graph, one shared text encoder, one shared VAE, identical seed and caption.

  • Base path: 25 steps, CFG 3.5
  • Distilled path: 16 steps, no CFG guidance
  • ImageCompare renders both next to the source

This is the graph that settles arguments. Before you commit a client deliverable to one checkpoint, run it here.

Note: the A/B graph ships with a 25-step base path rather than 50 — a deliberate time-saving default. Raise it if you want the full-fidelity comparison.

What Klein is actually good at (and what it is not)

Across the test runs, the strongest areas were:

Detail fidelity on complex subjects. Full-body figures, hands, shoes, handbags — pass rates were solid. Multi-element relationships (two people, a bicycle, a coffee cup) mostly held together.

Local edits. Earring swaps, cup-to-vase replacement, clothing recolours, standing-to-sitting pose changes — the majority landed cleanly on the tested subjects.

Multi-reference fusion. Character, outfit, and scene blending performed well within the 1–2 reference range.

The problems are real too:

  • Consistency drift on multi-reference work — the community reports it and I reproduced it
  • Occasional limb failures on complex poses
  • Bounding boxes are a strong hint, not a guarantee

None of these are unique to Klein. Qwen Image 2.1 shows the same failure modes. The difference is that Klein's floor is now high enough that these read as normal open-model limitations rather than dealbreakers.

Hardware notes

The distilled model on the INT8 build runs on 8 GB of VRAM, which is the headline number. In practice:

  • FP8 (5.6 GB) is the sensible default for consumer cards
  • BF16 (11.2 GB of weights) for quality-critical work on 24 GB+ cards
  • BFL reports roughly 4 seconds for a 1-megapixel image on an RTX 5090, and under 9 seconds on an RTX 4070 through ComfyUI (vendor figures, not re-measured here)

For local deployment, Diffusers added day-one support through ModularPipeline. One caveat worth knowing: the prompt-upsampler component can be popped out of the block graph if you are running a newer transformers, and for cover art or product shots with detailed prompts you will not miss it.

Getting the workflows

All seven graphs are published together, each with its own model entry, cover, and dependency list:

If you want to run these without a local GPU, the workflows are also set up on RunningHub:

Verdict

FLUX 3 Klein is not the model that beats everything. It is the model that makes "open weights, native 4K, editing, and multi-reference" a reasonable default instead of a compromise.

The distilled checkpoint is the one to use. The seven workflows cover the realistic use cases without demanding you understand the whole graph taxonomy first. And at 1.04 points behind Qwen Image 2.1 on a 100-image benchmark, the choice between them comes down to workflow preference rather than capability.

If your work involves local generation, 4K output, or editing that has to hold up on a deadline — this is worth testing this week.

Written after 28 comparison runs and seven workflow tests on locally deployed FLUX 3 Image Klein. Weights: black-forest-labs/flux-3-image-klein. Weights licensed under the FLUX Kommunity License.

Top comments (0)