ToolScore
Image Gen 3 min read

FLUX vs Qwen-Image on a 3060: Our Production Notes

Qwen for text, FLUX for look the number that matters here

By Stan K.

Two open models cover nearly all our image work on one RTX 3060 12GB in ComfyUI: FLUX.1-dev GGUF Q6_K (24 steps, cfg 1.0, guidance 3.5) for the look, Qwen-Image Q3_K_M (20 steps, cfg 2.5) for accurate text inside the image. Which one we reach for, the thumbnail LoRA — and the FLUX license line that trips people.

Two open-weight models cover nearly all of our image generation on a single RTX 3060 12GB in ComfyUI: FLUX.1-dev (GGUF Q6_K quant, a 9.86 GB file) and Qwen-Image (20B, GGUF Q3_K_M). They are not competitors. After months of production use they have settled into completely different jobs, and the split is cleaner than any benchmark table suggests.

The setup that fits in 12GB

ModelQuant we runOur settingsThe job it holds
FLUX.1-devGGUF Q6_K (9.86 GB file)24 steps, cfg 1.0, guidance 3.5overall look, aesthetics, mood
Qwen-ImageGGUF Q3_K_M (20B)20 steps, cfg 2.5text inside the image

Both run at in-pipeline quality — the same box renders for our video pipeline, as described in One GPU Beats Ten Subscriptions. One practical note we keep re-learning: Qwen’s text encoder is the heavy half of that model, and it spills into system RAM during sampling (ComfyUI handles this automatically). The 12GB of VRAM is enough, but starve the machine on system RAM and the whole pipeline feels it.

Qwen-Image: the one that can spell

The reason Qwen-Image keeps a permanent slot is a single capability: accurate text rendered inside the image. Titles, labels, signage, UI mockup copy — the things diffusion models spent a decade mangling — come out spelled correctly at our standard 20 steps and cfg 2.5. When the deliverable has words in it, we don’t audition models anymore.

FLUX: the one that can light a scene

FLUX.1-dev at 24 steps, cfg 1.0, guidance 3.5 remains our default for everything where the look is the product: photographic realism, composition, texture, mood. No open model we run on this card matches its general aesthetics — which is exactly why its license situation hurts (next section).

The thumbnail workflow

For YouTube thumbnails we run the youtube-thumbnails-flux LoRA on FLUX.1-dev — a trained dark, high-contrast thumbnail look that beats prompting your way there generically. The division of labor stays the same: FLUX plus the LoRA produces the image; when the thumbnail needs rendered-in text rather than an overlay, that render goes to Qwen.

NC model, commercial outputs — read the FLUX license twice

FLUX.1-dev’s license is non-commercial for the model, but its outputs “may be use[d] for any purpose (including for commercial purposes)”. You can run it and sell what it makes; what you cannot do is serve third parties with it — that needs a paid BFL license. Qwen-Image (the original) and Qwen-Image-Edit are Apache 2.0 with no such line. The full trap list — including the Qwen successor that quietly dropped those rights — is in The License Traps Hiding in “Open” Image Models.

When neither is the answer

The 3060 box is not the top of the quality ladder. When a job needs arena-topping quality (GPT Image 2.6 Sunburst sits at Elo 1262) or 4K output, we rent it per image instead of pretending a quantized 12GB rig competes — the crossover math and the honest break-even are in When Local Image Gen Beats Every API, and the full price table lives in the Image Gen hub. The hardware itself — what a 12GB card is worth in 2026 and what else it runs — is ranked in the Local GPU AI leaderboard.

Numbers change — the math doesn't
Token prices and plan limits shift monthly; the structural arguments in this guide age slowly. For current per-tool numbers see the live rankings.