Choose the model that supports the image you need to make.

List the finish, composition, and control requirements for the image. A model brief might ask for clean anime linework, stable adult character traits, a full-body pose reference, portrait dimensions, and a defined token ceiling. This list lets you reject incompatible options before generation.

Separate required features from preferences. Character reference support can be required for a recurring cast, while a soft painted finish may remain a preference. A model that lacks a required control leaves the workflow broken even if its sample images look attractive.

Lewdora shows active model information inside the generator because models and integrations can change. Check character reference, pose reference, image-to-image, dimensions, and other exposed settings. Do not rely on an old article or screenshot when the live interface presents a different capability set.

The model catalog also shows token cost before generation. Include compatible options in the calculation because hires processing or added controls can change the request cost. Compare the complete setup rather than the base model label.

Some anime models respond well to compact tags. Others read descriptive sentences with more precision. Start with the format shown in product guidance, then keep the content of the prompt stable across candidates. Rewriting the prompt for each model makes the comparison hard to interpret.

Use concrete visual language in either format: fictional adult subject, appearance anchors, action, frame, light, and finish. Avoid unsupported syntax copied from another platform. A weight or emphasis marker only helps when the current model and interface interpret it.

Choose two or three models that meet the brief. Keep prompt, dimensions, reference inputs, and comparable settings fixed. Generate several outputs per model and score prompt fit, composition, identity, anatomy, and finish against criteria written before the test.

Record missing or incompatible controls instead of forcing false parity. One model may support a reference mode that another lacks. That difference belongs in the conclusion because capability fit can matter more than a subjective preference between two selected images.

A low request cost can become expensive when you need many retries. Count the tokens spent on the full sample and divide by the outputs that meet your criteria. This accepted-result cost reflects prompt fit and repeat behavior better than the price of one generation.

Keep the sample size and quality bar consistent across models. A model should not receive credit for a selected success if its other outputs fail the same criteria. Use token pricing as one decision factor beside control support and visual fit.

Save the model label, prompt format, settings, reference behavior, cost, and known failure cases. Date the note because the active catalog can change. A small record prevents you from repeating the same comparison each time you return to a project.

Recheck the live catalog before a large batch. A model update or retired integration can change behavior. Treat published model advice as a workflow for evaluation, while the generator remains the source for current availability and cost.

List the controls and outputs the project actually needs before opening the catalog. Record required reference types, image-to-image support, dimensions, visual finish, number of characters, expected batch size, and token budget. Mark each requirement as essential or optional. A model that misses an essential control should leave the shortlist even if its sample style looks attractive.

Use the live capability record as the source of truth. Marketing labels cannot guarantee that character reference, pose reference, or a setting works on every model. The generator should expose compatible controls for the active choice. If a requirement is absent, treat it as unavailable until current product information documents otherwise.

Prepare two or three non-explicit prompts that represent the work: a close portrait, a full-body pose, and a scene with background interaction. Reuse the same visual brief, dimensions, and permitted references. Generate several outputs per model so one unusually strong or weak sample does not decide the result.

Score named criteria such as prompt adherence, identity retention, anatomy, composition, finish, and request cost. Keep the raw outputs and settings. Do not combine scores into a universal winner if the models serve different tasks. One may handle portraits well while another gives more reliable full-body structure or reference control.

Model catalogs evolve. Providers can add controls, retire versions, change pricing, or alter defaults. Date your comparison and record the exact model label shown during the run. Before starting a large series, repeat a small smoke sample with the saved evaluation packet to confirm that the expected behavior still holds.

Keep a fallback model for important work. The fallback should meet the essential controls even if its finish differs. Store a prompt adaptation for each model instead of assuming syntax transfers unchanged. This preparation reduces disruption when availability changes and makes the model decision an operational choice rather than a one-time aesthetic preference.

A lower request cost does not guarantee a lower project cost. Track how many attempts each model needs to produce an accepted image under the same brief. Multiply request cost by the observed attempts, then note time spent correcting prompts or references. Use this project-specific figure instead of a generic price claim.

Keep aesthetic preference separate from operational reliability. You may choose one model for a key illustration because its finish fits the art direction and another for a sequence because it retains identity more often. The catalog can support several roles without requiring one permanent winner.