Catalog
Which uncensored image model to use, and what each one is good at
Seedream 5 Pro produces the most striking work in the catalog, especially with reasoning switched on, while the Qwen 3 models are the more realistic choice for real people and the better editors for staying faithful to an original. Vellria Lite costs the least by a wide margin and is startlingly good at a narrow range of subjects.
Measured on September 8, 2026 against the catalog as it stood that day; models and their costs change.
How this was judged, and what it is not
This is a note about our own catalog, written from generating with these models rather than from reading about them. It compares nothing outside this platform. There is no scoring rubric behind it — it is judgement built from output, offered as judgement, and it says nothing about speed because generation time was never measured here.
What is exact is structural: the credit cost of each size, how many reference images an editing model accepts, and which parameters a model exposes. Those come from the catalog and appear on each model's own page rather than in this prose, because a number retyped into a sentence goes stale the day the catalog moves.
The catalog changes. This reflects it as of the date above.
Seedream 5 Pro, and the reasoning switch
Seedream 5 Pro is the strongest image model here and one of the strongest anywhere, both for general generation and for adult material. Its output is the most visually accomplished in the catalog, and it is the one to use when the image is meant to be looked at rather than merely correct.
It is the only model in the catalog that exposes a reasoning parameter, and turning it on changes the result substantially for the better. If you are using this model and not using that switch, you are leaving most of its advantage unused. It works with the switch off; it is a different class of model with it on.
The weakness is the flip side of the strength: it sometimes goes more artistic than you asked for. When the brief is plain and literal, that tendency works against you, and one of the Qwen models will land closer to what you described.
Qwen 3 and Qwen 3 Pro for realism and for editing
The two Qwen models are more realistic than Seedream, and the gap is widest on real people. Where Seedream reaches for an interpretation, Qwen renders the thing. For photographic work, or for anything where a face has to read as a person rather than as an illustration, these are the better starting point.
They are also the strongest editors in the catalog. Handed an existing image, they stay faithful to it — changing what you asked to change and leaving the rest where it was, which is the entire job of an editing model and the thing they most often get wrong.
Pro costs more than the base model and both share the same character. The base Qwen 3 is a sensible default; Pro is worth the difference when the result is going out rather than being checked.
Seedream 5 Lite and Wan 2.6 for the middle of the workload
Seedream 5 Lite is faster than Pro and still produces quality output, in both generation and editing. For ordinary work — the images that need to be good rather than exceptional — it is enough, and it costs less. It also accepts more reference images than any other editor in the catalog, which matters when a composite needs several sources.
Wan 2.6 splits sharply between its two modes. As an editor it is genuinely good and worth using. As a text-to-image model it is the weakest here: output reads as synthetic and consistency wavers in a way the others do not.
That split is worth taking literally. Reach for Wan 2.6 when you have an image to change, and reach past it when you are starting from a blank prompt.
Vellria Lite, and the language it wants
Vellria Lite is the cheapest model in the catalog by a wide margin and it is surprisingly good — not across the board, but on a specific range. Real people, particularly Asian subjects, and anime work come out far better than the price suggests they should.
It has one real limitation and it is worth knowing before you spend anything. It depends on the prompt language more than the others do: English and Chinese prompts work well, and other languages produce results that drift from what you asked for.
The dependency is a matter of degree rather than kind. Every model in the catalog does better with English or Chinese prompts; this one is simply the most sensitive to it. If your prompt is in another language and the result is not what you described, that is the first thing to change.
There is no editing counterpart. Vellria Lite generates from text only, so an image you want to modify has to go to one of the editors instead.
Picking one, in one paragraph
If the image needs to be striking, use Seedream 5 Pro with reasoning on. If it needs to look real, or if it is a photograph of a person, use Qwen 3 Pro. If you are editing an existing image, use Qwen 3 Pro or Seedream 5 Pro, or Seedream 5 Lite when the composite needs many references. If it is ordinary work, Seedream 5 Lite. If it is anime or an Asian subject and the budget is tight, Vellria Lite will surprise you. If you are starting from a blank prompt, do not use Wan 2.6.
Every model here runs the prompt as written, without a rewriting pass. That is a platform property rather than a model one, so it does not enter the comparison — but it is why the adult-content question above is about quality rather than about permission. The rules page covers what is refused, and the video model guide does this same exercise for clips.
Frequently asked questions
Should I always turn reasoning on for Seedream 5 Pro?
If quality is what you are after, yes — it is the single biggest improvement available on that model, and it is the only model in the catalog that offers it. Leave it off when you are checking whether a prompt is pointing in the right direction and intend to regenerate anyway.
Which model is most realistic for people?
The Qwen 3 models, and the gap over Seedream is clearest on real human subjects. Seedream tends toward interpretation, which is an advantage for striking work and a liability when the brief is a plain likeness.
Which editor stays closest to the original image?
Qwen 3 or Qwen 3 Pro. Both change what you asked to change and leave the rest alone. Seedream 5 Pro is also strong at editing, and Seedream 5 Lite accepts the most reference images of any editor here when a composite needs several sources.
Is Vellria Lite actually usable at that price?
On a narrow range, yes — real people, especially Asian subjects, and anime. Outside that range the more expensive models are better. Its real constraint is language: it depends on English or Chinese prompts more than the others do, and drifts on anything else.
Why avoid Wan 2.6 for text-to-image?
Because its output there reads as synthetic and consistency wavers, while every other model in the catalog does that job better. It is a different story as an editor, where it is genuinely good — the model is worth using, just not from a blank prompt.
Does prompt language really matter?
Yes, across the whole catalog. English and Chinese prompts produce better results everywhere here, and Vellria Lite is the most sensitive to it. If a result is drifting from your description and the prompt is in another language, rewrite it in English before changing anything else.