SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

Trying out NewBie image Exp0.1 in ComfyUI & Prompt Generation

まゆひらa

Last update 7-28-2026
* This is a model released in December 2025.
* Related articles are summarized in ComfyUI Article List.
* Please feel free to send your comments to the note Question Box . You might be happier sending messages other than questions. Please note that responses are whimsical.



■ 0. Overview

▼ 0-0. Introduction

　In this article, I will explain the steps to use "NewBie image Exp0.1" in ComfyUI. Also, I have tentatively devised a way to create XML structured prompts on AI chat services (free options available), which is unique to this architecture, and I will introduce it in section 4.

　Actually, I wrote an article back in December 2025 when the model was released, but it required a dedicated ComfyUI environment. Since ComfyUI provided native support that same month, I am following up with this article, albeit a bit late.

　Please see the article below for how to install ComfyUI.

▼ 0-1. About NewBie image Exp0.1

　Once again, I will list the features and model information. This is based on content output by ChatGPT.

  • Specialized for anime style

  • Next-DiT based architecture, different from SDXL's U-Net

  • The model is small, with 3.5B parameters, equivalent to SDXL

  • Text Encoder uses the small Gemma3-4B-it and Jina CLIP v2

  • VAE uses the one from FLUX.1

  • Prompts support XML structuring (see 3. to 4. for usage), which makes it easy to set multiple characters individually, as well as natural language and Danbooru tags

  • Basically supports 1MP (1024x1024, etc.) output. For reference, the official ComfyUI template is set to 1024x1536

  • Currently an experimental release

▼ 0-2. Reference Links

　These are links to the official project and model.

　These are official links.

This information is from before ComfyUI provided native support. The prompt writing style might be helpful. The sample images include workflows, allowing for prompt extraction.

The models used in this article (details below) and the official ComfyUI workflows, etc.


■ 1. Notes and Resource Usage

▼ 1-1. Notes (Standard Practices)

If you have not updated ComfyUI, necessary nodes or features may be missing. Since updates often include bug fixes, please keep it as up-to-date as possible.

Depending on the situation, models loaded into VRAM may be automatically offloaded to main RAM. Therefore, having more main RAM may allow for a more comfortable experience.

▼ 1-2. Required Resources and Generation Time

With a graphics card having 16GB of VRAM, there was not much headroom. Given the model size, it is thought that it would work without issues on 12GB if there is sufficient free main RAM.

Generation time per image on the author's environment (GeForce RTX 5060Ti 16GB). The first time includes the sum of both.

  • 5 seconds (model loading and text encoding)

  • 35 seconds (sampling only, 20 Steps, CFG=5.5, 1344x768)


■ 2. Model Installation

▼ 2-1. Comfy Org Version

　In this article, we will use the version distributed by Comfy Org. Download each file listed at the URL below and move them to the appropriate folders within your ComfyUI installation directory.

  • Comfy-Org/NewBie-image-Exp0.1_repackaged
    https://huggingface.co/Comfy-Org/NewBie-image-Exp0.1_repackaged

    • split_files/diffusion_models/
      NewBie-Image-Exp0.1-bf16.safetensors (6.49GB)
      → models\diffusion_models\

    • split_files/text_encoders/
      gemma_3_4b_it_bf16.safetensors (7.23GB)
      jina_clip_v2_bf16.safetensors (1.04GB)
      → models\text_encoders\

　Since the VAE is not listed above, we will download it from elsewhere.

　For reference, creating and separating folders like 'newbie' makes it easier to keep track of which architecture each file is for.


■ 3. Image Generation (for XML)

▼ 3-1. Overview

　I have posted the workflow used in ComfyUI and screenshots of the execution results.

　This article focuses on how to write XML-structured prompts.For prompts using natural language or Danbooru tags, please refer to 5. accordingly.

▼ 3-2. Workflow (Standard Version)

　I am providing the workflow for generation; please feel free to use it, including for modification and redistribution.

　This version has only one prompt field. The split version is posted in the next section.

(Last update 2026-7-27)

　Drag and drop the workflow file into the ComfyUI screen to load it, then click 'Run' to start generation. Images are saved in the 'output' folder of your ComfyUI installation.

　If the model locations or names differ, please adjust the settings accordingly. Feel free to change the prompts and resolution to your preference. The execution example below is the cover image for this article.

NewBie image Exp0.1 generation example
Image generated using the posted workflow

　For reference, I have posted the prompt used. You can see that it is a format not often seen for image generation.

You are an assistant designed to generate high-quality anime images with the highest degree of image-text alignment based on xml format textual prompts. <Prompt Start>
{
"character_1": {
  "bbox": [
    0,
    0,
    500,
    1000
  ],
  "name": "$character_1$"
},
"character_2": {
  "bbox": [
    500,
    0,
    1000,
    1000
  ],
  "name": "$character_2$"
},
"image": {
  "tags": "<character_1>
<n>$character_1$</n>
<gender>1girl</gender>
<appearance>young_girl, brown_hair, twin_tails, amber_eyes, fair_skin, round_face, rosy_cheeks</appearance>
<clothing>yellow_hoodie, denim_shorts</clothing>
<body_type>short, petite, childlike_build</body_type>
<expression>bright_smile, open_mouth, cheerful_expression, looking_at_character_2</expression>
<action>sitting_on_bench, talking, relaxed_pose</action>
<interaction>chatting_with_character_2, sitting_side_by_side, enjoying_conversation</interaction>
<position>center_left, foreground, medium_shot, zoomed_in</position>
</character_1>

<character_2>
<n>$character_2$</n>
<gender>1girl</gender>
<appearance>young_girl, blonde_hair, short_bob_hair, blue_eyes, fair_skin, soft_face, rosy_cheeks</appearance>
<clothing>pink_cardigan, white_dress</clothing>
<body_type>short, petite, childlike_build</body_type>
<expression>gentle_smile, happy_expression, looking_at_character_1</expression>
<action>sitting_on_bench, talking, relaxed_pose</action>
<interaction>chatting_with_character_1, sitting_side_by_side, enjoying_conversation</interaction>
<position>center_right, foreground, medium_shot, zoomed_in</position>
</character_2>

<general_tags>
<count>2girls, multiple_girls</count>
<style>anime_style, colorful_digital_art, detailed_shading, soft_lineart</style>
<background>park_bench, green_trees, soft_greenery, blurred_park_background</background>
<environment>quiet_public_park</environment>
<perspective>eye_level, medium_shot, zoomed_in, character_focused_composition</perspective>
<atmosphere>cheerful, warm, wholesome, peaceful</atmosphere>
<lighting>bright_daylight, soft_sunlight, gentle_shadows, warm_color_tone</lighting>
<objects>wooden_bench</objects>
<other>simple_composition, reduced_background_details, subtle_depth_of_field</other>
<quality>masterpiece, best_quality, highres, absurdres, detailed</quality>
</general_tags>",
  "caption": "An extremely detailed anime-style digital illustration features two young girls sitting side by side on a wooden bench in a quiet public park, shown in a zoomed-in medium shot with a character-focused composition. $character_1$ is positioned on the center-left in the foreground. She has brown twin tails, amber eyes, fair skin, rosy cheeks, and a round cheerful face. She wears a yellow hoodie and denim shorts. She sits in a relaxed pose, smiling brightly with an open-mouthed cheerful expression as she talks happily with $character_2$. $character_2$ is positioned on the center-right in the foreground. She has short blonde bobbed hair, blue eyes, fair skin, rosy cheeks, and a soft happy expression. She wears a pink cardigan over a white dress. She is also seated comfortably on the bench, turned slightly toward $character_1$ as they enjoy a lively and friendly conversation. The background is kept simple, with the bench surrounded by green trees and soft greenery in a lightly blurred park setting. The image is viewed from eye level, with unnecessary elements removed to keep attention on the girls’ expressions and interaction. Bright daylight, soft sunlight, and gentle shadows create a warm, peaceful, and wholesome atmosphere, rendered with detailed shading, soft lineart, and high-quality anime-style polish."
  }
}

▼ 3-3. Workflow (Bonus)

Compatible with the official ComfyUI workflow, this is a version with two prompt fields. Please feel free to use it, including for modifications and redistribution.

(Last update 2026-7-27)

Here, only the "{user_tags}" and "{user_caption}" sections can be set for the full prompt text below. Therefore, it is not possible to define drawing areas when multiple characters appear.

You are an assistant designed to generate high-quality anime images with the highest degree of image-text alignment based on xml format textual prompts. <Prompt Start>
{
"character_1": {
"bbox": [
0,
0,
1000,
1000
],
"name": "$character_1$"
},
"image": {"tags": "{user_tags}",
"caption": "{user_caption}"
}
}

This is the overall workflow and a generation example. Two input fields are provided at the bottom left, and the entire prompt is displayed at the bottom right during execution.

NewBie image Exp0.1 Generation Example 2
Image generated with the posted workflow 2

For reference, I have posted the contents of the tag field and the prompt field.

<character_1>
<n>$character_1$</n>
<gender>1girl</gender>
<appearance>pale_pink_hair, long_hair, ahoge, red_eyes, round_eyes, soft_face, youthful</appearance>
<clothing>large_hair_bow, stone_necklace, purple_wizard_dress, off_shoulder</clothing>
<expression>angry_smile, blush, open_mouth, cheerful_expression, looking_at_viewer</expression>
<action>raised_hands, lively_pose, upper_body_emphasis</action>
<position>upper_body, close_up, view_from_above</position>
</character_1>

<general_tags>
<count>1girl, solo</count>
<style>high_quality_shoujo_manga_style, premium_anime_illustration, soft_anime_style, pastel_colors, elegant_color_design, slightly_thicker_lineart, slightly_bold_lineart, ultra_refined_lineart, soft_shading, nuanced_shading, layered_shading, luminous_rendering, dreamy, cute</style>
<background>ruined_city, rubble, destroyed_buildings, small_fire_on_ground, starry_night_sky, simplified_background</background>
<environment>fantasy_ruins, ruined_city_at_night</environment>
<perspective>upper_body, close_up, view_from_above, character_focused_composition</perspective>
<atmosphere>magical, dreamy, whimsical, heartwarming, lively</atmosphere>
<lighting>moonlit_glow, magical_backlight, soft_dramatic_lighting, gentle_highlights, luminous_edge_lighting</lighting>
<quality>masterpiece, best_quality, ultra_highres, absurdres, extremely_detailed, refined_details, crisp_rendering, polished_finish, illustration_masterwork</quality>
<objects>magical_lightning_effect, floating_white_ghost, stuffed_ghost</objects>
<other>sparkling_effects, soft_glow, airy_feeling, shoujo_manga_tone, subtle_texture, fine_detail_emphasis, premium_finish, slightly_heavier_line_emphasis</other>
</general_tags>
An extremely detailed, high-quality shoujo manga style digital illustration features $character_1$ in a close-up upper-body composition viewed slightly from above. $character_1$ is a young girl with pale pink long hair, a playful ahoge, and round red eyes, giving her a sweet and expressive appearance. She wears a large hair bow, a stone necklace, and a purple off-shoulder wizard dress, creating a charming magical-girl impression. She shows an angry yet cheerful smiling expression with blush on her cheeks and an open mouth, looking directly at the viewer. Her upper body is emphasized as she raises her hands in a lively and expressive pose, drawing attention to her face, hair, and costume details. The scene is set against a softly simplified ruined city at night, with hints of rubble, destroyed buildings, small fires on the ground, and a starry sky. Magical lightning effects, a floating white ghost, and a stuffed ghost add playful fantasy elements around her. The illustration is rendered with slightly thicker and more noticeable lineart while preserving a refined shoujo manga appearance, along with elegant pastel color design, layered and nuanced shading, luminous highlights, subtle texture, and a polished premium finish. A gentle moonlit glow and magical backlighting create a whimsical, airy, and heartwarming atmosphere while maintaining a luxurious high-quality visual style.


■ 4. Generation of XML Structured Prompts

▼ 4-1. Overview

I believe XML structured prompts are difficult for humans to write directly. Therefore, I have prepared a manual for generating XML structured prompts on chats like Claude or ChatGPT.

The summary below uses content that ChatGPT added to and revised based on my proposal.

  • Simply attach the manual and enter the desired content to output an appropriate prompt. Measures have also been taken to prevent accidental image generation.

  • Output in a code block to make the prompt easy to copy.

  • Designed to grasp the intent and output a somewhat detailed prompt even if the input is short.

  • Organized and output by category, such as character, background, composition, lighting, and atmosphere.

  • Ensured that unknown information is not forced and that the tag and caption contents do not contradict each other.

  • Designed not to include author tags unless instructed.

▼ 4-2. Prompt Creation Manual

　This was created mainly using ChatGPT. Please feel free to refer to it, modify it, etc. Usage instructions are explained in the next section.

　The following two will be output in a format that can be used with the standard version workflow in 3-2. (the full text of the prompt to be entered).

【Basic Version】
　Usually, please choose this one. The "character_n": part at the beginning of the prompt is fixed, but the official ComfyUI workflow uses a similar mechanism.

（Last update 2026-7-27）

【Strict Version】
　Here, the "character_n": part at the beginning of the prompt can be adjusted. This allows you to define the drawing area when multiple characters appear. It is recommended to use this after understanding the specifications of the XML structured prompt.

（Last update 2026-7-27）

【Bonus Version】
　This will be output in a format that can be used with the workflow in 3-3. or the official ComfyUI workflow. Since the contents of "tags:" and "caption:" are output individually, enter them into their respective fields in the workflow.

（Last update 2026-7-27）

▼ 4-2. Prompt Generation 1

　Attach any of the files from the previous section to a new chat in ChatGPT, Claude, Gemini, Grok, etc. and enter the image you want to create.

　In this section, I will explain an example using Claude. Even if you do not have an account yet, you can easily sign up if you have a Google account.

　At the time of writing, the free version of Claude has "Sonnet 3.5 + Thinking" selected by default. You can reduce usage consumption by setting the effort to low and turning off thinking.

Free version of Claude screen (excerpt)

　If you drag and drop any of the files from the previous section, it will be attached, so enter your preferred content and press Enter.

　In the manual, I do not have it decide things that have not been instructed. Therefore, by entering it so that the phrase "specifically" applies to the whole thing, the prompt is slightly expanded.

　Here, I will attach "NewBie-image-Exp0.1_Prompt_Creation_Manual_Basic.txt" and enter "Specifically devise a scene showing a girl in casual, cute clothes standing by a lake, from the waist up."

Attached the file and entered the input

　Once the response is returned, you can move the mouse cursor to the upper right corner of the prompt to copy it.

Copy the output prompt

　Finally, paste it into the prompt field of the workflow to generate the image.

I generated an image using a prompt created with Claude.
Generated image

For reference, I have also included the workflow that can generate the image above.

▼ 4-2. Prompt Generation 2

Since the manual focuses on prompt writing rules (prompt creation instructions are just a bonus), you can also input image prompts or the images themselves to generate prompts. However, please be aware of the possibility that reliance on the input may occur.

In this section, I will input an image to obtain a prompt. I attached the image and the manual, and entered, "Convert the content of the image into a prompt." Perhaps because of the student-style clothing, Claude refused to output the prompt, so I used Grok (https://grok.com/) instead.

Prompt output by Grok

The image generated using the prompt output by Grok.

Generated image

The input image is the menu cover from June 2026 (from Miscellaneous Notes 2026). While not identical, you can see that the content is quite similar. I believe that the ability to finely control the layout and image content using only prompts is a major feature of this architecture.

Image input to obtain the prompt


■ 5. For natural language and Danbooru tags

▼ 5-1. Overview

Although this article is written assuming XML structured prompts, standard natural language and tag formats can also be used. Similar to 4., I have prepared a manual for generating prompts on chats like Claude or ChatGPT. Please refer to 4. for the overview and usage of the manual.

▼ 5-2. Prompt Creation Manual

Similar to 4-2., this was created mainly using ChatGPT.

(Last update 2026-7-27)

This is mostly a bonus, so I will omit usage examples. Basically, only the format of the output prompt changes.


■ 6. Finally

▼ 6-1. From the author

　The December '25 article focused mainly on the generation procedure, so I couldn't delve into how to create prompts. It has been over half a year, but I believe I have been able to properly follow up on that point.

　It is an architecture with personality and potential, so I am sincerely hoping for the arrival of a new version.


■ 7. Others

　Please follow the menu to see other articles I have written.

　There is also a list of articles limited to ComfyUI.

　If you have any questions regarding the articles, please contact @riddi0908 on X.

いいなと思ったら応援しよう！