这些设定图是怎么做出来的。共 3 步,中间那一步是故意会失败的。 1. 先在 Midjourney 里把角色捏出来。先把脸调到满意,再考虑设定图的事。 2. 把那张渲染图喂给 GPT Imag...

GPT-Image-2 Iancu_ai Sun Sep 06 14:25:17 +0000 2026

How the sheets were made. 3 steps, and the middle one fails on purpose.

1. Build the character in Midjourney first. Get the face you want before you think about the sheet.

2. Feed that render into GPT Image 2 to lay the sheet out. This is the step that breaks. You get the layout you asked for and a face that has drifted off your Midjourney character.

The layout half of that prompt, verbatim:

LEFT HALF, two full body figures standing side by side, small in frame, head to toe visible: the same costume seen from the front and from the back. Straight neutral stance, arms hanging at the sides, feet together. Both figures are CUT OFF AT THE NECK, no head, no face, no hair on these two figures, the body simply ends at a clean neckline like a headless mannequin.

RIGHT HALF, one single large head and shoulders portrait of the character, facing the camera straight on, neutral expression, lit evenly on the same white background. This is the only face anywhere in the image.

Absolutely no text, no letters, no numbers, no labels, no logos, no watermarks. Exactly one face in the whole image. The two body figures must be headless.

3. Put the face back with a 2 reference redraw. Image 1 is the sheet you just got, image 2 is your Midjourney render:

Reproduce IMAGE 1 exactly, with one single change: replace the head in the large portrait on the right with the head from IMAGE 2. Everything else must stay identical to IMAGE 1, the pure white background, the layout, the two headless full body figures on the left and every detail of their costume, the framing, the lighting, the scale. Do not redesign the costume. Do not add a head to the two body figures, they stay cut off at the neck.

That alone gives you a head sitting straight on the collar with no neck. You have to say this out loud:

ANATOMY, the most important requirement: the head must sit on a real neck. Leave a clear visible stretch of bare skin between the jawline and the top of the collar, roughly the height of the chin. The collar sits low, at the base of the neck near the collarbones, NOT against the jaw. You may lower the top edge of the collar to make room; keep everything else about the costume identical. The neck reads as a proper cylinder with its own shading and the tendons visible. Do not let the chin rest on the armour. Do not make the head look pasted on. Pull the portrait framing back a little so the head, the whole neck and the shoulders are all in frame.

The permission to lower the collar is the part that does the work. Without it the model can only slide the head upward, and that looks worse than the problem.

这些设定图是怎么做出来的。共 3 步,中间那一步是故意会失败的。 1. 先在 Midjourney 里把角色捏出来。先把脸调到满意,再考虑设定图的事。 2. 把那张渲染图喂给 GPT Image 2,让它负责排版。这一步会出问题。你会得到你想要的排版,但脸上的样子已经偏离你原本在 Midjourney 里捏的角色了。 那段提示词里负责排版的那一半,原样如下: LEFT HALF, two full body figures standing side by side, small in frame, head to toe visible: the same costume seen from the front and from the back. Straight neutral stance, arms...

一键生图查看来源

中文说明

这些设定图是怎么做出来的。共 3 步,中间那一步是故意会失败的。 1. 先在 Midjourney 里把角色捏出来。先把脸调到满意,再考虑设定图的事。 2. 把那张渲染图喂给 GPT Image 2,让它负责排版。这一步会出问题。你会得到你想要的排版,但脸上的样子已经偏离你原本在 Midjourney 里捏的角色了。那段提示词里负责排版的那一半,原样如下: LEFT HALF, two full body figures standing side by side, small in frame, head to toe visible: the same costume seen from the front and from the back. Straight neutral stance, arms...

原始 Prompt

How the sheets were made. 3 steps, and the middle one fails on purpose.

1. Build the character in Midjourney first. Get the face you want before you think about the sheet.

2. Feed that render into GPT Image 2 to lay the sheet out. This is the step that breaks. You get the layout you asked for and a face that has drifted off your Midjourney character.

The layout half of that prompt, verbatim:

LEFT HALF, two full body figures standing side by side, small in frame, head to toe visible: the same costume seen from the front and from the back. Straight neutral stance, arms hanging at the sides, feet together. Both figures are CUT OFF AT THE NECK, no head, no face, no hair on these two figures, the body simply ends at a clean neckline like a headless mannequin.

RIGHT HALF, one single large head and shoulders portrait of the character, facing the camera straight on, neutral expression, lit evenly on the same white background. This is the only face anywhere in the image.

Absolutely no text, no letters, no numbers, no labels, no logos, no watermarks. Exactly one face in the whole image. The two body figures must be headless.

3. Put the face back with a 2 reference redraw. Image 1 is the sheet you just got, image 2 is your Midjourney render:

Reproduce IMAGE 1 exactly, with one single change: replace the head in the large portrait on the right with the head from IMAGE 2. Everything else must stay identical to IMAGE 1, the pure white background, the layout, the two headless full body figures on the left and every detail of their costume, the framing, the lighting, the scale. Do not redesign the costume. Do not add a head to the two body figures, they stay cut off at the neck.

That alone gives you a head sitting straight on the collar with no neck. You have to say this out loud:

ANATOMY, the most important requirement: the head must sit on a real neck. Leave a clear visible stretch of bare skin between the jawline and the top of the collar, roughly the height of the chin. The collar sits low, at the base of the neck near the collarbones, NOT against the jaw. You may lower the top edge of the collar to make room; keep everything else about the costume identical. The neck reads as a proper cylinder with its own shading and the tendons visible. Do not let the chin rest on the armour. Do not make the head look pasted on. Pull the portrait framing back a little so the head, the whole neck and the shoulders are all in frame.

The permission to lower the collar is the part that does the work. Without it the model can only slide the head upward, and that looks worse than the problem.

近景特写多分屏长发多面板布局角色人物肖像概念设计 3D 渲染