Qwen-Image-2.1 vs Bernini-Diffusers-v2: The Licence Is the Whole Argument

Two open-weight image systems landed six weeks apart in 2026 with almost no fanfare, and they disagree about the one thing that decides whether you can build a business on them. Qwen-Image-2.1, published on 20 September 2026, ships under a research licence — non-commercial only, commercial use by separate arrangement. Bernini-Diffusers-v2, published by ByteDance on 13 August 2026, ships under Apache 2.0, which means you can ship it. The more capable-looking release is the one you cannot sell, and the more permissive one is the one that mostly wants to make video. That inversion is the comparison.

The licence inversion

Start with the paperwork, because it outranks everything else on this page. No amount of quality gap closes a licensing problem, and no benchmark score rescues a product that has to be pulled.

• Licence — Qwen-Image-2.1 is Qwen Research License Agreement, dated 20 September 2026, "FOR NON-COMMERCIAL PURPOSES ONLY", with commercial rights available only on request; Bernini-Diffusers-v2 is Apache 2.0.
• Practical consequence — a Qwen-Image-2.1 workflow is an evaluation asset or a personal one; a Bernini-Diffusers-v2 workflow is shippable today, with the usual Apache attribution requirements and no vendor conversation needed.
• Direction of travel — Qwen's earlier image line shipped under Apache 2.0, so this release is a tightening; ByteDance has kept Bernini permissive across both generations.

If you only read one section of this article, that is the section. Everything below is a tiebreaker between two things that are not actually tied on this axis.

What each system actually is

The two models are not the same kind of object, and the difference shows up the moment you look at what ships in the download.

Qwen-Image-2.1 is a single unified model. One 7B, 32-layer single-stream diffusion transformer handles generation and editing in the same weights, conditioned by a Qwen3-VL 8B text encoder, with a 64-channel RGBA VAE at 16× spatial compression doing the heavy lifting on transparency. Around 33 GB total, with the DiT itself at roughly 14 GB and the text encoder taking the rest. It also ships two optional prompt-rewriting checkpoints — PE-T2I and PE-I2I, each a fine-tuned Qwen3.5-VL 9B at about 18.8 GB — which rewrite your instruction into something precise before the diffusion model sees it.

Bernini-Diffusers-v2 is a pipeline, not a single model. A Qwen2.5-VL semantic planner does the reasoning and writes a latent semantic plan; a Wan2.2 diffusion renderer executes it. The v2 recipe's headline change over the first Bernini release is that it warms up the connector for thousands of steps before co-training, which is what improved reference-guided video editing and OpenS2V behaviour. That two-stage split is the architectural bet: plan first, render second, and you can edit video by editing the plan.

• Architecture — Qwen-Image-2.1 is one unified DiT for generation and editing; Bernini-Diffusers-v2 is a Qwen2.5-VL planner driving a Wan2.2 renderer.
• Modality — Qwen-Image-2.1 is images only; Bernini-Diffusers-v2 covers t2i, i2i, t2v, v2v, reference-guided video editing and r2v, so images are a side room rather than the house.
• Transparency — Qwen-Image-2.1 produces alpha out of the sampler through its RGBA VAE; Bernini-Diffusers-v2's documented output is RGB video and image frames, with no equivalent native alpha path described.
• Prompt handling — Qwen-Image-2.1 ships separate 9B rewriter checkpoints you can read, diff and override; Bernini-Diffusers-v2's planner is part of the pipeline and its prompt enhancer is an optional OpenAI-compatible endpoint call rather than a shipped checkpoint.
• Hardware — Qwen-Image-2.1 fits a single consumer GPU with CPU offload recommended on constrained cards; Bernini-Diffusers-v2's documented environment is Hopper-class, H100/H800/H200, with multi-GPU sequence parallelism for the larger configurations.

Where the comparison stops being useful

It is tempting to score these two against each other on output quality. That comparison does not exist, and pretending otherwise would be the dishonest move.

Bernini-Diffusers-v2's published figures are vendor-reported video and editing metrics — EditVerse 8.02, OpenVE 3.96, OpenS2V 63.83, VBench 84.46 — plus a claim of first-tier performance in ByteDance's internal blind human pairwise arena. They are video metrics. No independent image benchmark of Bernini-Diffusers-v2 has been run, no FID, no text-rendering score, no image leaderboard entry. On the other side, Qwen-Image-2.1's only quality number is the vendor's own Qwen-Image-Bench chart, where it reads 60.28 against Nano Banana 2.0 at 59.82 and GPT Image 1.5 at 59.65 — vendor material, unreproduced, and the model was absent from the independent image leaderboards when this was written. The one genuinely independent data point is a functional verification from the SGLang integration work: transparent PNG output passed, alpha was retained across the full 0–255 range, RGBA PSNR measured 60.69 dB in a single sample, and FP8 configurations passed transparent-PNG generation. The SGLang authors are explicit that this is a correctness check, not a quality guarantee.

So: two open releases, two vendor scorecards, no shared benchmark, and no third party who has run both. Anyone who tells you one beats the other on image quality is repeating marketing.

Editing philosophy: rewrite the prompt, or rewrite the plan

Both models are trying to solve the same underlying problem — a user's instruction is vague and the diffusion model is literal — and they solve it in opposite places.

Qwen-Image-2.1 solves it in language. The PE-I2I checkpoint takes your editing instruction plus one or more input images and rewrites it into a precise directive before generation. Because it is a shipped checkpoint with a shipped system_prompt.txt, you can read exactly what it is told to do — including the rule that rendered text in the output should match the dominant language of text already in the input image, rather than being forced to English. That is inspectable behaviour. You can override it.

Bernini-Diffusers-v2 solves it in structure. The planner produces a semantic plan that the renderer then executes, and the interesting property of that arrangement is that the plan is a thing you can edit. For video work — where the whole point is often to change one element across a sequence without re-rolling the shot — planning separately from rendering is the right shape. For a single still, it is an extra moving part.

• Editing granularity — Qwen-Image-2.1 supports local edits by circle, painted annotation or a separate mask, plus transparent-layer editing and subject extraction from RGB photographs; Bernini-Diffusers-v2's documented editing strengths are reference-guided video editing and subject-to-video, not still-image local repair.
• Reference inputs — Qwen-Image-2.1 accepts up to 10 reference images, though an early-access reviewer reported multi-reference consistency degrading from about three images onward; Bernini-Diffusers-v2's reference handling is documented around video sequences.
• Inspectability — Qwen-Image-2.1's rewriting logic is a readable system prompt in the repository; Bernini-Diffusers-v2's planning stage is model weights.

What to actually do with this

The decision rule falls out of the licence rather than the spec sheet.

If the output is going into something you sell — client work, a product, a paid pipeline — Bernini-Diffusers-v2 is the only one of these two you can use without a conversation, and you should budget for Hopper-class hardware to run it as documented. If you are doing research, evaluation, or personal work, Qwen-Image-2.1 gives you the more inspectable system and the one with native alpha, at the cost of a licence you will eventually have to renegotiate or walk away from.

There is a third option that neither of these covers, and it is worth naming plainly: hosted image models, where you never touch a licence file or a GPU. OrcaRouter routes 200+ models behind one OpenAI-compatible endpoint at provider list price with zero markup, with automatic failover across providers, a routing DSL for fallback chains, and model fusion. Neither Qwen-Image-2.1 nor Bernini-Diffusers-v2 is in our catalogue at the time of writing — we route no Qwen-Image model of any version and no Bernini model — so this is not a pitch for either. The routed image line we do carry is OpenAI's GPT-Image family, Google's Imagen 4 tiers and Gemini image previews, and xAI's Grok Imagine image endpoint; because list price is passed through rather than marked up, a vendor price cut on any of them is live the same day.

The part neither repository answers

Both of these were quiet releases with no run-up, and quiet releases are hard to plan around. Qwen-Image-2.1's only advance signal was a 17 September early-access note offering 50 ModelScope slots with testers asked to publish by 29 September — three days before the weights went public. Bernini-Diffusers-v2 simply appeared on Hugging Face on 13 August with a Gitee mirror alongside it.

What is missing from both is a reason to believe the terms will hold. Qwen tightened its licence between the previous image generation and this one, which is a data point about direction. ByteDance has kept Bernini Apache 2.0 across two generations, which is a data point about the other direction. Whether either of those holds for the next release is unknowable from here, and it is the thing to check before you commit a pipeline rather than after.