Skip to content

[Issue]: Sprite generation prompt alteration #2871

Description

@Stegorix

Summary

When generating a sprite or sprite sheet the final prompt is missing information and edited preview prompts do
not correspond to the sent prompt.

Expected behavior

The prompt does not miss critical information and the edited preview is sent to the image api as is

Actual behavior

See steps to reproduce

Steps to reproduce

(With enabled prompt preview)

  1. Start sprite generation and set Description (For example to "cat")
  2. "Generate Sheet"
  3. See tag "cat" is missing from the review (Samples under "Logs, screenshots, or video")
  4. Clear prompt (and negative prompt) and only add "Cat"
  5. "Generate"
  6. Image prompt to the api includes infos that where deleted from the preview before

Environment

  • Marinara Engine version: 2.0.5
  • Install type: docker
  • Browser or app shell: Vivaldi

Logs, screenshots, or video

Generation with Appearance Description Cat and no edits (1 (1x1) Neutral) (No prompt preview enabled):

Prompt from Marinara: Preview not enabled, so not visible

Prompt sent to API:

prompt: "detailed eyes, facial expression: neutral, solo, transparent background, visual novel sprite, clean lineart, full body, single character portrait sprite, head and shoulders portrait, centered in frame, anime style, illustration, best quality

MANDATORY SPRITE SHEET LAYOUT: return one 512x512px image containing exactly 1 separate cells in a strict 1 columns by 1 rows grid. Each cell is exactly 512x512px; vertical grid cuts are every 512px and horizontal grid cuts are every 512px. Fill every cell. The first 1 cells, read left-to-right then top-to-bottom, must be: neutral. No missing cells, no extra cells, no merged cells, no blank cells, no uneven grid, and no one-large-image composition. Do not return a 3x3 grid, 9 cells, or fewer cells than requested.",
  negative_prompt: "cropping, grid, panel borders, text, labels, watermark, photorealistic, 3d render, lowres, bad anatomy, bad hands, logo, signature, missing cells, fewer than 1 cells, extra cells, merged cells, blank cells, uneven grid, one large image spanning cells, 3x3 grid, 9 cells" 

Info Cat missing

Generation with Appearance Description Cat and no edits (1 (1x1) Neutral):

Prompt from Marinara:

prompt: "detailed eyes, facial expression: neutral, solo, transparent background, visual novel sprite, clean lineart, full body, single character portrait sprite, head and shoulders portrait, centered in frame, anime style, illustration, best quality

MANDATORY SPRITE SHEET LAYOUT: return one 512x512px image containing exactly 1 separate cells in a strict 1 columns by 1 rows grid. Each cell is exactly 512x512px; vertical grid cuts are every 512px and horizontal grid cuts are every 512px. Fill every cell. The first 1 cells, read left-to-right then top-to-bottom, must be: neutral. No missing cells, no extra cells, no merged cells, no blank cells, no uneven grid, and no one-large-image composition. Do not return a 3x3 grid, 9 cells, or fewer cells than requested."
negative prompt: "cropping, grid, panel borders, text, labels, watermark, photorealistic, 3d render, lowres, bad anatomy, bad hands, logo, signature, missing cells, fewer than 1 cells, extra cells, merged cells, blank cells, uneven grid, one large image spanning cells, 3x3 grid, 9 cells" 

Prompt sent to API:

prompt: "detailed eyes, facial expression: neutral, solo, transparent background, visual novel sprite, clean lineart, full body, single character portrait sprite, head and shoulders portrait, centered in frame, anime style, illustration, best quality

MANDATORY SPRITE SHEET LAYOUT: return one 512x512px image containing exactly 1 separate cells in a strict 1 columns by 1 rows grid. Each cell is exactly 512x512px; vertical grid cuts are every 512px and horizontal grid cuts are every 512px. Fill every cell. The first 1 cells, read left-to-right then top-to-bottom, must be: neutral. No missing cells, no extra cells, no merged cells, no blank cells, no uneven grid, and no one-large-image composition. Do not return a 3x3 grid, 9 cells, or fewer cells than requested.

MANDATORY SPRITE SHEET LAYOUT: return one 512x512px image containing exactly 1 separate cells in a strict 1 columns by 1 rows grid. Each cell is exactly 512x512px; vertical grid cuts are every 512px and horizontal grid cuts are every 512px. Fill every cell. The first 1 cells, read left-to-right then top-to-bottom, must be: neutral. No missing cells, no extra cells, no merged cells, no blank cells, no uneven grid, and no one-large-image composition. Do not return a 3x3 grid, 9 cells, or fewer cells than requested.",
  negative_prompt: "cropping, grid, panel borders, text, labels, watermark, photorealistic, 3d render, lowres, bad anatomy, bad hands, logo, signature, missing cells, fewer than 1 cells, extra cells, merged cells, blank cells, uneven grid, one large image spanning cells, 3x3 grid, 9 cells, missing cells, fewer than 1 cells, extra cells, merged cells, blank cells, uneven grid, one large image spanning cells, 3x3 grid, 9 cells", 

Marinara added MANDATORY SPRITE SHEET LAYOUT and negative prompt again before sending to API

Generation with Appearance Description Cat and clear preview and set to "dog" (1 (1x1) Neutral):

Prompt from Marinara: Manually edited to

prompt: "dog"
negative_prompt: "" 

Prompt sent to API:

prompt: "dog

MANDATORY SPRITE SHEET LAYOUT: return one 512x512px image containing exactly 1 separate cells in a strict 1 columns by 1 rows grid. Each cell is exactly 512x512px; vertical grid cuts are every 512px and horizontal grid cuts are every 512px. Fill every cell. The first 1 cells, read left-to-right then top-to-bottom, must be: neutral. No missing cells, no extra cells, no merged cells, no blank cells, no uneven grid, and no one-large-image composition. Do not return a 3x3 grid, 9 cells, or fewer cells than requested.",
  negative_prompt: "missing cells, fewer than 1 cells, extra cells, merged cells, blank cells, uneven grid, one large image spanning cells, 3x3 grid, 9 cells", 

Marinara added MANDATORY SPRITE SHEET LAYOUT and negative prompt again before sending to API

Additional context

More infos:

  • The missing "cat" tag seems to be influenced by the Style profiles/Prompt grammar, but i was not able to find a setting that influences the sprite prompts
  • (Unrelated) The "MANDATORY SPRITE SHEET LAYOUT" part seems a bit too much for most local image generation models: In testing Z-Image handled them "Ok" (Grid with expressions, but too many and not in requested order), while SDXL generates an empty grid at best.

Template check

  • I DID NOT read this template and provide the requested details.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions