Flux.3 Klein Morning Briefing 2026-11-14 | Top Pick: FLUX 3 Image Klein 12GB — 5.6B int8 Text to Image + Edit with up to 4 Reference Images + B

This issue reviews 1 new versions of Flux.3 Klein, listed in Column A. Top pick: The_frizzy1's 12GB workflow: performs text-to-image generation and editing with up to 4 reference images using the official int8 model on a 12 GB GPU, with BFL's official prompt upsampler enabled by default. The author provides itemized speed tests and peak VRAM usage. Since it is essentially the BFL official workflow with added comments and low-VRAM defaults, without new weights, it is listed in Column A rather than S; there are no downgrades or exclusions this issue.

Zhenzhen's AI Radar · Flux.3 Klein · 2026-11-14 Morning Briefing

This issue features 1 new versions: Strongly Recommended 0 · Recommended 1 · Worth Following 0 · Included Only 0

Editor's Note This issue reviews 1 new versions of Flux.3 Klein, listed in Column A. Top pick: The_frizzy1's 12GB workflow: performs text-to-image generation and editing with up to 4 reference images using the official int8 model on a 12 GB GPU, with BFL's official prompt upsampler enabled by default. The author provides itemized speed tests and peak VRAM usage. Since it is essentially the BFL official workflow with added comments and low-VRAM defaults, without new weights, it is listed in Column A rather than S; there are no downgrades or exclusions this issue.

Recommended (1)

[A] FLUX 3 Image Klein 12GB — 5.6B int8 Text to Image + Edit with up to 4 Reference Images + BFL Prompt Upsampler · v1.0.0

Workflow · New Model · Author The_frizzy1 · Published 11-14 05:39 · Downloads 12

Runs two workflows for Flux.3 Klein (5.6B distilled version) on a 12 GB GPU: text-to-image (with optional bounding boxes) and editing with up to 4 reference images, with BFL's official prompt upsampler enabled by default.

Images

Example 1024x1024

Reasons for Recommendation

  • New workflow based on the BFL official int8 workflow; the author added comments and set the default resolution to 1 MP. Uses only core nodes, requires ComfyUI 0.41 or higher.
  • The author explains why the upsampler remains enabled: BFL's model card states that Klein was trained on long scene descriptions, so short prompts yield mediocre results; the upsampler model (Qwen3.5 9B) examines the input image first in edit mode.
  • The author tested it on an RTX 3060 with 12 GB VRAM, reporting peak VRAM usage of 11.4 GB; 6 visible examples provided.

How to Use

  • Required 4 files: the int8 main model, the Qwen3-VL 4B text encoder, the prompt expansion model, and the flux2 VAE; you must first log in to the Hugging Face repository and agree to the BFL terms
  • The 3th and 4th reference images in the edit are muted by default; select them and press Ctrl+M to enable them. Only disable the enable upsampler option when writing long descriptions yourself; the expanded result can be viewed in the Upsampled prompt field
  • The author did not specify the number of sampling steps separately, so the official workflow settings are used

Risks and Limitations

  • The author states that 4 MP runs on 12 GB VRAM, but color banding appeared in the sky of test images; stopping at 1 MP is recommended.
  • FLUX Kommunity License: The model is non-commercial; commercial use of outputs is limited to annual revenues under $500 thousand. Read the license before selling.
  • It is essentially the official workflow with added comments; new content mainly consists of low-VRAM defaults and explanations; 3 examples are hidden.

Data: Civitai public API (base model Flux.3 Klein), inclusion window 11-13 20:31 → 11-14 09:18 (Beijing Time). Reasons for recommendation have been reviewed by editors; download counts reflect attention only.

Civitai

Explore the latest AI Radar updates