Gemma 4 models leaking "think" capability

[61150n4n1m35-svg]

I would like to report a possible bug or issue regarding how ComfyUI loads the Gemma 4 model. Recent updates have changed the way Gemma 4 models are loaded; up to version 0.25, the model worked perfectly as an LLM for vision tasks—such as describing images or refining/expanding prompts—outputting only the description or the expanded prompt. However, starting with version 0.29, the behavior changed: the model now "leaks" its thought process, acting more like a general assistant and ignoring system prompt instructions by including its internal reasoning in the output, even when the system prompt explicitly forbids this. In the images below, I show the two versions tested in Stability Matrix, along with screenshots of the results: the correct description from version 0.25 versus the description containing <think> tags from version 0.30. Both tests used the same workflow (from ComfyUI's built-in templates), the recommended Gemma 4 model, and identical prompts and images, yet yielded vastly different results.

Tested versions

Result from ComfyUI current version 0.30

Result from ComfyUI version 0.25

Here I am comparing the two results generated in the same workflow across two different versions of ComfyUI.

Result in ComfyUI version 0.25
This is a vibrant and energetic piece of fan art featuring the character Hitori "Bocchi" Gotoh from the anime and manga series Bocchi the Rock!.

Here is a detailed description:

Character:

• The central figure is Bocchi, depicted in a dynamic pose. She has long, flowing, bright pink hair with bangs.
• She is dressed in an outfit that suggests a rock musician aesthetic: a dark (likely black) t-shirt featuring a graphic design, and a short, flowing blue skirt that seems to be tied or draped around her hips.
• She is holding an electric guitar, which is black and appears heavily stylized, suggesting a powerful instrument. Her posture is active, as if she is mid-performance or about to strike a chord.
• Her legs are visible, showing long, slender limbs, and she is wearing dark socks with red shoelaces tied around her ankles, paired with dark shoes.
• Her expression is intense, fitting the high energy of the surrounding text.

Composition and Background:

• The background is dominated by bold, distressed typography in shades of red and white, reading "BOCCHI THE "GUITAR HER..." and "BOCHI ROCK!". The text is large, fragmented, and gives the impression of a concert poster or album cover.
• The lower portion of the image features a textured, earthy brown and tan ground, giving the scene a gritty, stage-like atmosphere.
• The overall color palette is high-contrast, combining the bright pink of her hair and the vibrant reds of the text against the darker tones of her clothing and the background.

Mood and Style:

• The art style is highly polished, characteristic of modern anime illustration—clean lines, expressive features, and dramatic lighting.
• The mood is energetic, passionate, and rebellious, perfectly capturing the spirit of rock music and the character's journey toward musical confidence.

Result in ComfyUI version 0.30 (current)
Here's a thinking process to arrive at the detailed image description:

1. Analyze the Request: The user wants a description of the provided image.

2. First Impression & Subject Identification:

   • What is the main focus? An anime character.
   • Who is the character? She strongly resembles Hitori "Bocchi" Gotoh from the anime Bocchi the Rock!.
   • What is she doing? She is posing dynamically while holding an electric guitar.
   • What is the overall mood/style? Energetic, rock-and-roll, vibrant (despite the dark background).

3. Detailed Breakdown - The Character:

   • Appearance: Long, bright pink hair. Large, expressive blue eyes. Fair complexion.
   • Attire: She is wearing a dark (likely black or navy) t-shirt with some graphic design on it. She has a short, flowing skirt, which appears to be a mix of darker and possibly blue/purple tones. She is wearing dark socks and heavy, dark shoes with red laces.
   • Pose: Dynamic and energetic. She is leaning slightly, one leg extended forward, suggesting movement or performance. Her arms are engaged with the guitar.
   • Prop: She is holding a black electric guitar, which looks like a standard rock instrument.

4. Detailed Breakdown - The Background & Text:

   • Background: Dark, earthy brown/sepia tones. It looks textured, almost like rough ground or a stage backdrop, giving it a gritty, rock aesthetic.
   • Text/Branding: Prominently displayed in large, stylized, distressed red text across the top right is "BOCCHI!" followed by more obscured text that reads "...ROCK!". There is also smaller text near the top left mentioning "BOCCHI THE 'GUITAR HER...". This confirms the source material.

5. Synthesis and Structuring the Description: A good description moves from general to specific.

   • Introduction (General Overview): Identify the style, subject, and theme immediately. (Anime illustration, rock music theme, Bocchi).
   • Mid-Section (Character Focus): Describe the character's appearance, clothing, and action (the core of the image).
   • Final Section (Setting & Atmosphere): Describe the background, color palette, and any significant text elements that set the mood.

6. Refining the Language (Self-Correction/Enhancement):

   • Instead of just saying "pink hair," use descriptive words like "long," "vibrant," or "pastel."
   • Instead of "dark background," use "gritty," "earthy," or "textured brown."
   • Ensure the tone matches the energy of the image (dynamic, intense).

7. Final Polish: (This leads directly to the polished response provided below.)This is a vibrant, dynamic illustration in the style of modern anime, heavily themed around rock music. The central figure is a character strongly resembling Hitori "Bocchi" Gotoh from the series Bocchi the Rock!.

The Character:
The character has long, voluminous, bright pink hair and large, expressive blue eyes. She is depicted in a highly energetic pose, mid-performance. She is dressed in a casual but edgy outfit: a dark (likely black) t-shirt featuring a graphic design, paired with a short, flowing skirt that incorporates shades of dark blue and purple. She wears dark socks and sturdy, dark shoes laced with red string.

She is actively playing or posing with a sleek, black electric guitar, which is held across her body. Her posture is confident and lively, conveying the energy of a live musician.

The Background and Atmosphere:
The background is dominated by deep, earthy brown and sepia tones, giving the scene a gritty, raw, and atmospheric feel, reminiscent of a dimly lit concert venue or a worn stage.

Overlaying the top portion of the image is bold, distressed typography in bright red. Large, impactful lettering spells out "BOCCHI!" and fragments of other text, such as "...ROCK!", clearly branding the artwork to the musical theme of the character.

Overall Impression:
The image successfully captures the blend of shy vulnerability (implied by the character's usual demeanor) and explosive musical passion, using strong color contrast (bright pink and blue against the dark browns and red text) to create a high-energy, rock-and-roll aesthetic.

[blepping]

You used a different prompt in the two examples. Describe the image, for 0.25 and Describe the image (without a trailing comma) for 0.3. That can affect the response.

[61150n4n1m35-svg]

A new test was run—this time using the same prompt for both—and the results are identical; the presence of a comma in one of the prompts made no significant difference. The fact is that the new version of ComfyUI is leaking the model's "thought process" along with the requested output.
Here are the new tests:
ComfyUI 0.25

comfyui 0.30

Full prompt result:
ComfyUI 0.25
This is a vibrant, stylized illustration featuring the character Hitori "Bocchi" Gotoh from the anime and manga series Bocchi the Rock!.

Here is a detailed description:

Character:

• Appearance: Bocchi is depicted as a young female character with long, voluminous, light pink hair that flows around her. She has large, expressive blue eyes.
• Attire: She is dressed in a dark, possibly black or navy, short-sleeved top (which appears to have some graphic design on it) paired with a flowing, deep blue skirt that falls to mid-thigh. On her feet, she wears dark socks and chunky, dark-colored sneakers with red accents on the laces.
• Pose: She is striking a dynamic, energetic pose, leaning slightly forward with one leg extended and the other bent, giving her a sense of movement and performance.

Action and Props:

• She is holding an electric guitar, which is a prominent feature. The guitar is black and appears to be plugged in, suggesting she is actively playing or about to play. Her hands are positioned on the neck and body of the instrument.

Background and Atmosphere:

• The background is dark and dramatic, dominated by shades of brown and black, suggesting a gritty stage or concert environment.
• Large, bold, distressed typography frames the character. The word "BOCCI" is visible in massive, textured red lettering on the left, and the word "ROCK!" is visible in similar style on the right. This text strongly reinforces the musical theme of the artwork.
• There are small, subtle white circular elements scattered in the background, perhaps representing stage lights or energy effects.

Overall Impression:
The illustration is highly energetic and captures the essence of rock music. The contrast between Bocchi's bright pink hair and the dark, intense background creates a visually striking and passionate portrayal of a musician.

Comfyui 0.30
Here's a thinking process to arrive at the detailed image description:

1. Analyze the Request: The user wants a description of the provided image.

2. First Impression & Subject Identification:

   • What is the main focus? An anime character.
   • Who is the character? She strongly resembles Hitori "Bocchi" Gotoh from the anime Bocchi the Rock!.
   • What is she doing? She is posing dynamically while holding an electric guitar.
   • What is the overall mood/style? Energetic, rock-and-roll, vibrant (despite the dark background).

3. Detailed Breakdown - The Character:

   • Appearance: Long, bright pink hair. Large, expressive blue eyes. Fair complexion.
   • Attire: She is wearing a dark (likely black or navy) t-shirt with some graphic design on it. She has a short, flowing skirt, which appears to be a mix of darker and possibly blue/purple tones. She is wearing dark socks and heavy, dark shoes with red laces.
   • Pose: Dynamic and energetic. She is leaning slightly, one leg extended forward, suggesting movement or performance. Her arms are engaged with the guitar.
   • Prop: She is holding a black electric guitar, which looks like a standard rock instrument.

4. Detailed Breakdown - The Background & Text:

   • Background: Dark, earthy brown/sepia tones. It looks textured, almost like rough ground or a stage backdrop, giving it a gritty, rock aesthetic.
   • Text/Branding: Prominently displayed in large, stylized, distressed red text across the top right is "BOCCHI!" followed by more obscured text that reads "...ROCK!". There is also smaller text near the top left mentioning "BOCCHI THE 'GUITAR HER...". This confirms the source material.

5. Synthesis and Structuring the Description: A good description moves from general to specific.

   • Introduction (General Overview): Identify the style, subject, and theme immediately. (Anime illustration, rock music theme, Bocchi).
   • Mid-Section (Character Focus): Describe the character's appearance, clothing, and action (the core of the image).
   • Final Section (Setting & Atmosphere): Describe the background, color palette, and any significant text elements that set the mood.

6. Refining the Language (Self-Correction/Enhancement):

   • Instead of just saying "pink hair," use descriptive words like "long," "vibrant," or "pastel."
   • Instead of "dark background," use "gritty," "earthy," or "textured brown."
   • Ensure the tone matches the energy of the image (dynamic, intense).

7. Final Polish: (This leads directly to the polished response provided below.)This is a vibrant, dynamic illustration in the style of modern anime, heavily themed around rock music. The central figure is a character strongly resembling Hitori "Bocchi" Gotoh from the series Bocchi the Rock!.

The Character:
The character has long, voluminous, bright pink hair and large, expressive blue eyes. She is depicted in a highly energetic pose, mid-performance. She is dressed in a casual but edgy outfit: a dark (likely black) t-shirt featuring a graphic design, paired with a short, flowing skirt that incorporates shades of dark blue and purple. She wears dark socks and sturdy, dark shoes laced with red string.

She is actively playing or posing with a sleek, black electric guitar, which is held across her body. Her posture is confident and lively, conveying the energy of a live musician.

The Background and Atmosphere:
The background is dominated by deep, earthy brown and sepia tones, giving the scene a gritty, raw, and atmospheric feel, reminiscent of a dimly lit concert venue or a worn stage.

Overlaying the top portion of the image is bold, distressed typography in bright red. Large, impactful lettering spells out "BOCCHI!" and fragments of other text, such as "...ROCK!", clearly branding the artwork to the musical theme of the character.

Overall Impression:
The image successfully captures the blend of shy vulnerability (implied by the character's usual demeanor) and explosive musical passion, using strong color contrast (bright pink and blue against the dark browns and red text) to create a high-energy, rock-and-roll aesthetic.

I also ran an additional test in ComfyUI 0.30 by adding two rules to the prompt, but it still leaks the "think" process at the beginning of the generated output; whereas in version 0.25, no instructions are needed—simply asking for a description without any extra rules results in an image description free of any internal model thought processes.

Result of the ComfyUI 0.30 test adding rules: (only describe), (no thinking)

Full result of the prompt generated in ComfyUI 0.30 with rules:
The user wants me to describe the provided image.
The image is an anime-style illustration of a young girl playing an electric guitar.

Description elements:

1. Subject: A young, pink-haired girl.
2. Action/Pose: She is holding and playing a black electric guitar, striking a dynamic pose.
3. Attire: She is wearing a dark t-shirt (with a graphic), a blue skirt, and dark socks with red shoelaces on her shoes.
4. Background: The background is dark brown/black and features large, stylized red text that reads "BOCCHI THE 'GUITAR HERO' ROCK!".
5. Style: Anime/manga illustration style.

I will synthesize these points into a concise description, adhering strictly to the "only describe" instruction.An anime-style illustration features a young girl with long, flowing pink hair. She is dressed in a dark t-shirt, a blue skirt, and dark socks with red laces, and she is actively playing a black electric guitar. The background is dark brown and features large, stylized red text that reads "BOCCHI THE 'GUITAR HERO' ROCK!".

In this test, you can clearly see the model start to hallucinate when rules are added to prevent its "thought process" from leaking; it thinks in the first line, but then mixes part of that thought with the rest of the image description. In contrast, with earlier versions of ComfyUI (like 0.25), a simple prompt without any rules yields the correct description immediately, with no extra clutter.

I tried other ways to prevent the model's "thought" process from leaking into the output, but nothing worked: reducing the token count just results in an incomplete thought, and the description doesn't fit within fewer than 2024 tokens; lowering the temperature only affects the description, while the thought process still appears in the result.

If there is another way to work around this issue, I would appreciate it if someone could share a solution—or perhaps the ComfyUI author will include a fix in the next update.