# Stable Diffusion 3.5 Variants, Licensing and Prompting

> Published 2026-08-30 · https://www.promptzone.com/zuri_wang/stable-diffusion-35-variants-licensing-and-prompting-1e15

Stable Diffusion 3.5 is worth understanding for two reasons that have nothing to do with leaderboards: its license permits commercial use for small organisations, and it kept real classifier-free guidance, so negative prompts still work. This covers what each variant in the family is for, what the license actually says, and how prompting differs from the FLUX habits many people have picked up since.

## Where 3.5 sits

Stability AI released Stable Diffusion 3 Medium in June 2024 to a rough reception, mostly over human anatomy, and the open-weights conversation moved to Black Forest Labs and FLUX almost immediately afterwards. Stable Diffusion 3.5 arrived in October 2024 as the correction: Large and Large Turbo first, with a Medium checkpoint following shortly after, all published on [Hugging Face](https://huggingface.co/stabilityai/stable-diffusion-3.5-large).

Architecturally the family is a multimodal diffusion transformer conditioned by two CLIP text encoders plus a T5 encoder. The practical consequence of that combination shows up in the prompting section below: the model responds to both keyword-style and sentence-style prompts, which is unusual.

| Variant | Rough size | Steps | Guidance | Suited to |
| --- | --- | --- | --- | --- |
| 3.5 Large | 8B class | 20-40 | real CFG, around 3.5-4.5 | final images, best prompt adherence |
| 3.5 Large Turbo | 8B class, distilled | a handful | keep CFG at or near 1 | drafting, high-volume iteration |
| 3.5 Medium | 2.5B class | 20-40 | real CFG | consumer GPUs, fine-tuning experiments |

Medium is the one to look at if your card is modest. It is a meaningfully smaller model and it runs where the 8B checkpoints will not, at some cost in composition and detail.

## The license is the real differentiator

For anyone building a product this matters more than image quality. Stable Diffusion 3.5 ships under the Stability AI Community License, which permits research, non-commercial use, and commercial use by individuals and organisations below an annual revenue threshold that Stability sets at one million US dollars. Above that line you need an enterprise agreement.

Compare that to the alternatives in the same weight class. FLUX.1 [dev] weights are open but the license is non-commercial, full stop. FLUX.1 [schnell] is Apache 2.0 and unrestricted, but it is a heavily distilled model. Stable Diffusion 3.5 occupies the gap: an undistilled model you can legally ship with while you are small.

Read the license yourself rather than trusting a summary, including this one. The threshold, the attribution requirement and the terms on outputs all matter, and licenses get revised.

![A printed contract and reading glasses on a wooden desk](https://cdn.stocksnap.io/img-thumbs/960w/1B8NHMDSDR.jpg)

## Negative prompts still work here

This is the other structural difference and it gets overlooked. FLUX [dev] and [schnell] have guidance distilled into the model, so the negative prompt field in your interface does nothing. Stable Diffusion 3.5 Large and Medium run genuine classifier-free guidance, which means two things:

- A negative prompt is functional. `blurry, watermark, extra fingers, text` behaves as it did on [SDXL](/jaroslav/how-to-install-and-run-sdxl-models-in-comfyui-a-complete-guide-2nk2).
- The CFG scale behaves the way you remember: too low and composition goes soft, too high and colours burn. The 3.5-4.5 band is a sensible starting range.

Large Turbo is the exception. It is distilled for few-step sampling and wants CFG at or near 1, exactly like FLUX schnell.

If your workflow depends on subtractive control, or on the ControlNet and inpainting ecosystem built around real guidance, this is a concrete reason to keep a 3.5 checkpoint installed alongside whatever else you use.

![A spread of vintage fashion magazine covers on a table](https://images.rawpixel.com/editor_1024/cHJpdmF0ZS9sci9pbWFnZXMvd2Vic2l0ZS8yMDIzLTA3L3BkbWlzYzE1LWxvYzIwMTQ2NDkxNDAtaW1hZ2UuanBn.jpg)

## Prompting: keywords and prose both work

Because 3.5 conditions on CLIP encoders as well as T5, both prompt dialects land. Keyword stacks inherited from the SD 1.5 and SDXL era still function, and full sentences with bound clauses also function. That flexibility is convenient and it makes one style of prompt particularly effective: a comma-separated stack where each item names a distinct, verifiable attribute.

Here is a compact example that blends two genres:

```plaintext
1960s glamour shot, a zombie in a fashion shoot wearing a hippie tunic with ethnic prints, 1960s bohemian style, raw flesh, peeling skin, decaying, film grain, american highway in the background, desaturated colors, 1960s magazine cover
```

It runs on Stable Diffusion 3.5 Large at CFG around 4, and it also works on FLUX.1 [dev] and on SDXL checkpoints, which is a useful property in a prompt you plan to reuse.

The structure is worth copying:

1. **Declare the genre first.** `1960s glamour shot` sets lighting, pose vocabulary and framing before anything else is specified.
2. **State the collision.** A zombie at a fashion shoot is the whole idea. Put it early and state it plainly.
3. **Give the collision physical detail.** `raw flesh`, `peeling skin`, `decaying` stop the model from producing a person in makeup.
4. **Anchor the era twice.** The decade appears in the genre, the styling and the output format. Repetition across different attributes is what makes a period read convincingly.
5. **Constrain the palette and the medium.** `desaturated colors`, `film grain`, `magazine cover` tell the model what the image physically is, not just what it depicts.

That last point is the transferable one. Naming the artefact, a magazine cover, a contact sheet, a lookbook page, does more for coherence than any quality adjective.

## Known weak spots

The 3.5 family renders text reasonably but not reliably, and complex hand positions still fail at a normal rate. Multi-subject scenes with distinct described attributes bleed into one another more than they do on the larger closed models. None of that is unusual for open weights in this class.

The more common practical problem is expectation transfer. If you have been running FLUX, your prompts are probably prose-heavy and guidance-light, and dropping them into 3.5 unchanged with CFG at 3.5 gives underwhelming results. Rebuild the prompt in the model's own dialect before judging it.

![An empty two-lane highway crossing desert scrub at dusk](https://cdn.stocksnap.io/img-thumbs/960w/FYRJ3CSO45.jpg)

## Takeaways

- Pick Large for final work, Large Turbo for drafting at low CFG, Medium when VRAM is the constraint.
- The Community License permits commercial use below Stability's revenue threshold, which makes 3.5 a legal option where FLUX [dev] is not. Read the current text before you rely on it.
- Negative prompts and real CFG work on Large and Medium. That alone justifies keeping a checkpoint around.
- Keyword stacks work well: genre, collision, physical detail, doubled era anchor, palette, artefact type.
- Do not port FLUX prompts unchanged. The dialects differ and so do the guidance ranges.

## Related reading

- [Filename Prompts: Making FLUX Output Look Like Real Photos](/nadim_nasrallah/filename-prompts-making-flux-output-look-like-real-photos-25o0)
- [FLUX 1.1 Pro and When a Closed Image Model Earns Its Cost](/carmen_jung/flux-11-pro-and-when-a-closed-image-model-earns-its-cost-ol3)
- [Writing Cinematic Portrait Prompts for FLUX Image Models](/ishaan_kobayashi/writing-cinematic-portrait-prompts-for-flux-image-models-12k5)