The model picker

In short

• Never pick a model and then get refused when you press Generate
• Open the MODELS tab in the Studio’s right panel
• It opens on USABLE: the models that can actually do what you have set up. ALL MODELS is one press away (Usable, or All Models)
• Every grayed card tells you why it is grayed, and the model you picked is never hidden from you (Why a card is grayed)
• Each model has its own prompt length limit. Go over it and you are stopped before anything is charged, and the PROMPT AGENT will rewrite it to fit (Prompt limits)
• Attach a picture that shifts the others and the Studio warns you, then renumbers the prompt to match (When positions move)

Usable, or All Models

The Studio’s model panel has a two-way switch at the top. Usable is the default and lists only the models that can do the operation you have chosen, at the duration you have asked for. All Models lists everything, with each unusable card grayed and carrying the reason on its face.

The count beside the title says which you are looking at: 12 models, or 7 of 12 models.

Three rules, and each one exists because the alternative lies to you:

The selected model is never hidden. A panel that hid your active choice would misrepresent what Generate is about to send. If the model you have picked is unusable for what you have asked, its card stays, grayed, with its reason.

The list is recomputed wherever anything can change it. Edit the duration, change the operation, switch mode, pick a different model: the list re-lists in the same breath. There is no reload and no second list to fall out of step.

The panel is never empty. When Usable would leave you with only the selected card, or nothing, the panel says so: No model here is usable — the rest are grayed for the reason on each card, with the switch to All Models right beside it.

Which view you are on is remembered by your browser, not stored on the prompt. It is what you are looking at, so it does not mark the prompt as changed.

Why a card is grayed

There is exactly one definition of unusable, and the card always says which of these it hit:

Reason	Reads roughly
No reference path	Lorvara has no reference path wired for X — it sends this model no reference images
No text-to-video	X has no text-to-video endpoint — it needs an input it does not offer
Frames	X takes a start frame but no end frame
No edit endpoint	X has no edit endpoint — it cannot edit an image
Retired	X is retired — it is no longer offered
Too short	Caps at 5s — 5s shorter than this 10s shot

The first five are facts about what Lorvara has wired to that model, said as such rather than dressed up as a fact about the model itself. A model may well accept references somewhere; what the card tells you is whether this product sends them.

The last one is worth watching on a shot list. A model whose longest clip is shorter than the shot you wrote will not stretch. It will cut your shot short, and it is grayed rather than allowed to do that quietly.

Prompt limits

Every model has a hard limit on its prompt field, and it is a model property, not a Lorvara one. The spread is large: the tightest model on the roster takes a thousand characters and the most generous takes sixty thousand. The whole Kling family sits at 2,500.

Two consequences follow, and both are worth knowing before they bite:

A prompt is written for one model’s limit. Prompts composed from a shot list are written to fit the project’s default model for that mode. Switch the prompt to a tighter model and it can be refused before it is sent, purely on length. The request is stopped and nothing is charged. That is not a fault to route around: it is the default doing what a default is for. The Prompt Agent will refactor a prompt to a new model’s limit; ask it.

A line break counts as two. The limit is counted in a unit where a newline costs two characters, so a prompt broken into paragraphs is longer than it looks. This is measured, not assumed: it is how a 2,487-character prompt came to be refused by a model whose stated limit is 2,500. The agent is told about it, and it is why a prompt that reads comfortably short can still be over.

What you write is not all that is sent. Lorvara appends the reference tokens in the model’s own spelling and a line naming what each attached picture is. The room the agent is given is the limit with that measured and taken off, so it writes to a budget that accounts for it.

When the reference positions move

A composed prompt ends with a line stating the order its references arrive in: Reference images, in the order sent: …. Those positions were true when it was composed.

Attach a picture afterwards and it can be sent first, pushing every reference after it down one. The prompt’s text then points at the wrong pictures, and nothing about the count gives it away: the count went up.

So the Studio compares the mapping, not the sizes: each picture the stored line names, found by what it is, against where it now sits. When they disagree you get a warning, The reference positions have moved, naming what moved where, with Renumber to match beside it.

The repair renumbers the tokens in the text and replaces the order line. Your words are not touched, and it is only offered when every token in the text belongs to a picture that is still being sent. When one does not, the warning stands on its own and the judgment is yours.

A picture the line names that cannot be identified unambiguously is never guessed at.

What to read next

• How do Models work? — why there is a roster at all.
• Seedance — the two versions, and what each accepts.
• How a prompt is assembled from a shot — where the limit is spent.
• How we’re tracking models — how the roster is kept current.