Vellria Flash
The fastest and cheapest image model. For drafts, variations and quick tests.
Text to image1 credit
Art workflow
An AI art workflow: draft, keep the frame, refine it
Almost nothing good comes out of a single prompt, and the catalog is arranged around that: draft cheaply until a frame is worth keeping, then hand that frame back to an editing model as a reference so the next pass refines what already works instead of starting the idea over. Because every model states its own cost before you commit and the balance is prepaid rather than metered by a plan, the arithmetic of a long series is available while the composition is still open.
The loop, not the single prompt
Almost nothing good comes out of one attempt. You run a spread of cheap drafts to find out what the idea looks like outside your head, discard most of them, and keep the one where the pose or the light or the staging is right. That frame is the real starting point. Everything after it is refinement, which is a different job from invention: you are protecting what already works while changing one part of it.
The catalog is arranged so that split costs you something to ignore. Some image models are cheap enough to run over and over while the composition is still open; others are dearer and meant for the frame you intend to deliver. Drafting on the expensive one and finishing on the cheap one is the usual way to spend a budget on nothing, and it is an easy habit to fall into, because the dearer model is more pleasant to be wrong on.
Models
Seedream 5 Pro
The flagship image model, focused on detail and composition. 2K output only.
Text to image9 credits
Seedream 5 Pro Edit
The flagship editing model. Takes up to ten references and preserves identity and lighting.
Image to image9 credits
Qwen 3 Pro
Strong on long, complex prompts; stays faithful to scene composition and text.
Text to image6 credits
Qwen 3 Pro Edit
Edits using the images you provide as references, preserving character and scene.
Image to image6 credits
Seedream 5 Lite
Fast, cheap Seedream. Reaches 4K and is good at typography and poster work.
Text to image6 credits
Seedream 5 Lite Edit
Takes up to fourteen references and redraws the images you provide.
Image to image6 credits
A prompt describes; a reference asserts
A prompt is a description, and descriptions get interpreted. It is the only part of a text-only request that varies between attempts, so when a result disappoints, the wording is where the answer has to be: a lower camera, a harder key light, rain on the street instead of after it. Changing one thing at a time is how you find out which word did the work.
A reference image asserts instead. It carries what you cannot reliably spell out, a face, a body type, one specific jacket, the exact grade of a light you already found, and hands it over as fact rather than as a sentence waiting to be interpreted. There is no adjective for a particular jaw. The editing models in the catalog take your own picture as that reference, and several take a set, so you can pin the character with one image and the setting with another.
That is what makes the loop close. The output of a draft is an input to the next pass, so the work accumulates instead of restarting, and your good frames become raw material rather than finished items you either keep or lose.
Holding a character across a series
Consistency is the first thing that breaks when every image starts from text. Two runs of the same description give you two people who merely resemble each other, and no amount of adjectives closes the gap, because you are re-negotiating the face on every attempt. The structural fix is to stop describing the character after the first success and start supplying it.
Generate the subject once, keep that file, and feed it back as the reference on every later image in the set. The character stops being a description and becomes an input. The same move works for a location, a costume, a colour grade. And when a later frame drifts, replace the reference rather than rewriting the prompt, because drift means the evidence was wrong, not the wording.
This is why the anchor frame belongs on your own disk rather than in your account. The platform holds finished files for a while, not forever, and a reference you no longer have is a series you can no longer continue; the page on how generation works gives the window.
Knowing where the line is before you build on it
A refusal that arrives late is expensive in a way its error message never shows. An anatomy study, a horror plate, a weapon in a period scene: a filter tuned for a mass consumer product treats a wide band of ordinary art as borderline, and it tends to say so after you have already built the composition around the subject and spent a pass framing it.
The refused render is not the real cost. The real cost is that you start writing around the filter, softer nouns, avoided subjects, an idea trimmed to keep the tool cooperative, and over the length of a project that becomes a self-censorship you stop noticing you are performing. You end up with the work the tool would allow rather than the work you set out to make, and there is no diff to show you the difference.
Your wording is not rewritten here, so what comes back is what you asked for, and what we prohibit outright is short, absolute and published on the content rules page rather than left to be inferred from an error message. The provider that renders a model keeps a filter of its own on top of that list, judged on its terms rather than ours, and it can stop a run after it has started; when it does, the credits taken at the start come back automatically, so what that layer costs a series is a re-run rather than a rule you have to guess at. A line you can read at the start of a project is one you can plan a series around; a fence you find in the dark is one you rebuild around, late and at your own expense.
Material from outside, and where the file goes after
Nothing in the loop requires the picture to have started here. Your own photograph, a scan of a drawing, a render out of another tool: they all attach the same way as something you generated, which matters if you are bringing an existing character, a real location or a piece of physical work into a series.
The finished file goes out the same way. Generation is one stage in a longer chain, not the whole of it, and the parts a generator is bad at, exact typography, precise compositing, print preparation, are the parts your existing tools are good at. Take the frame, take it into whatever you already use, and treat the model as a source of material rather than as the place the work has to finish.
Frequently asked questions
What is the difference between changing the prompt and adding a reference image?
The prompt is for what changes; the reference is for what has to stay. Wording moves the camera, the light, the action. A reference carries a face, a body type or one specific piece of wardrobe as fact, so the model reproduces it instead of reinterpreting your description of it on every run.
How do I keep the same character across a whole series?
Generate the character once, keep that frame, and use it as the reference on every later image rather than describing the person again. The editing models take your own picture, and several take a set, so you can pin the character with one image and the location with another. When a frame drifts, swap the reference rather than rewriting the description.
Can I work from my own photographs or scans?
Yes. A photograph, a scan of a drawing or a render from another tool attaches exactly as something generated here would, so an existing character or a real place can come into the series without being described from scratch. The editing models are the ones that read it.
Which model should I draft on and which should I finish on?
Draft on a cheap one and finish on a dear one. The catalog spans a wide range of costs, and each model page states its own before you commit, so the arithmetic is available while the composition is still open. Reserve the expensive run for the frame you intend to deliver.
Are the subjects a consumer tool refuses allowed here?
Many of them, yes. Fictional adult content is permitted, and so is the wider band of subject matter a mass-market filter treats as borderline. What is prohibited is short and absolute: anyone who is not of age, and real people used without their consent. The content rules have the full policy.