Quick-start
- Upload or drop an image in the left-hand panel.
- Pick a Caption Type and, if you wish, adjust the Caption Length.
- (Optional) expand the "Extra Options" accordion and tick any boxes that should influence the caption.
- (Optional) open Generation settings to adjust
temperature, top-p, or max tokens.
- Press Caption.
The prompt sent to the model appears in the Prompt box (editable),
and the resulting caption streams into the Caption box.
Caption Types
| Mode | What it does |
|---|
| Descriptive |
Formal, detailed prose description. |
| Descriptive (Casual) |
Similar to Descriptive but with a friendlier, conversational tone. |
| Straightforward |
Objective, no fluff, and more succinct than Descriptive. |
| Stable Diffusion Prompt |
Reverse-engineers a prompt that could have produced the image in a SD/T2I model.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| MidJourney |
Same idea as above but tuned to MidJourney’s prompt style.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| Danbooru tag list |
Comma-separated tags strictly following Danbooru conventions
(artist:, copyright:, etc.). Lower-case underscores only.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| e621 tag list |
Alphabetical, namespaced tags in e621 style – includes species/meta
tags when relevant.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| Rule34 tag list |
Rule34 style alphabetical tag dump; artist/copyright/character
prefixes first.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| Booru-like tag list |
Looser tag list when you want labels but not a specific Booru format.
⚠︎ Experimental – can glitch ≈ 3% of the time. |
| Art Critic |
Paragraph of art-historical commentary: composition, symbolism, style,
lighting, movement, etc. |
| Product Listing |
Short marketing copy as if selling the depicted object. |
| Social Media Post |
Catchy caption aimed at platforms like Instagram or BlueSky. |
Note on Booru modes: They’re tuned for
anime-style / illustration imagery; accuracy drops on real-world photographs
or highly abstract artwork.
Extra Options
These check-boxes fine-tune what the model should or should not mention:
lighting, camera angle, aesthetic rating, profanity, etc.
Toggle them before hitting Caption; the prompt box will update
instantly.
Generation settings
- Temperature – randomness.
0 = deterministic; higher = more variety.
- Top-p – nucleus sampling cutoff. Lower = safer,
higher = freer.
- Max New Tokens – hard stop for the model’s output length.
Enjoy experimenting, and feel free to open an issue if you
spot any bugs or have feature ideas!
🚨🚨🚨 If the "Help improve JoyCaption" box is checked, the _text_ query you write will be logged and I _might_ use it to help improve JoyCaption.
It does not log images, user data, etc; only the text query. I cannot see what images you send, and frankly, I don't want to. But knowing what kinds of instructions
and queries users want JoyCaption to handle will help guide me in building JoyCaption's dataset. This dataset will be made public. As always, the model itself is completely
public and free to use outside of this space. And, of course, I have no control nor access to what HuggingFace, which are graciously hosting this space, collects.
"""
CAPTION_TYPE_MAP = {
"Descriptive": [
"Write a detailed description for this image.",
"Write a detailed description for this image in {word_count} words or less.",
"Write a {length} detailed description for this image.",
],
"Descriptive (Casual)": [
"Write a descriptive caption for this image in a casual tone.",
"Write a descriptive caption for this image in a casual tone within {word_count} words.",
"Write a {length} descriptive caption for this image in a casual tone.",
],
"Straightforward": [
"Write a straightforward caption for this image. Begin with the main subject and medium. Mention pivotal elements—people, objects, scenery—using confident, definite language. Focus on concrete details like color, shape, texture, and spatial relationships. Show how elements interact. Omit mood and speculative wording. If text is present, quote it exactly. Note any watermarks, signatures, or compression artifacts. Never mention what's absent, resolution, or unobservable details. Vary your sentence structure and keep the description concise, without starting with “This image is…” or similar phrasing.",
"Write a straightforward caption for this image within {word_count} words. Begin with the main subject and medium. Mention pivotal elements—people, objects, scenery—using confident, definite language. Focus on concrete details like color, shape, texture, and spatial relationships. Show how elements interact. Omit mood and speculative wording. If text is present, quote it exactly. Note any watermarks, signatures, or compression artifacts. Never mention what's absent, resolution, or unobservable details. Vary your sentence structure and keep the description concise, without starting with “This image is…” or similar phrasing.",
"Write a {length} straightforward caption for this image. Begin with the main subject and medium. Mention pivotal elements—people, objects, scenery—using confident, definite language. Focus on concrete details like color, shape, texture, and spatial relationships. Show how elements interact. Omit mood and speculative wording. If text is present, quote it exactly. Note any watermarks, signatures, or compression artifacts. Never mention what's absent, resolution, or unobservable details. Vary your sentence structure and keep the description concise, without starting with “This image is…” or similar phrasing.",
],
"Stable Diffusion Prompt": [
"Output a stable diffusion prompt that is indistinguishable from a real stable diffusion prompt.",
"Output a stable diffusion prompt that is indistinguishable from a real stable diffusion prompt. {word_count} words or less.",
"Output a {length} stable diffusion prompt that is indistinguishable from a real stable diffusion prompt.",
],
"MidJourney": [
"Write a MidJourney prompt for this image.",
"Write a MidJourney prompt for this image within {word_count} words.",
"Write a {length} MidJourney prompt for this image.",
],
"Danbooru tag list": [
"Generate only comma-separated Danbooru tags (lowercase_underscores). Strict order: `artist:`, `copyright:`, `character:`, `meta:`, then general tags. Include counts (1girl), appearance, clothing, accessories, pose, expression, actions, background. Use precise Danbooru syntax. No extra text.",
"Generate only comma-separated Danbooru tags (lowercase_underscores). Strict order: `artist:`, `copyright:`, `character:`, `meta:`, then general tags. Include counts (1girl), appearance, clothing, accessories, pose, expression, actions, background. Use precise Danbooru syntax. No extra text. {word_count} words or less.",
"Generate only comma-separated Danbooru tags (lowercase_underscores). Strict order: `artist:`, `copyright:`, `character:`, `meta:`, then general tags. Include counts (1girl), appearance, clothing, accessories, pose, expression, actions, background. Use precise Danbooru syntax. No extra text. {length} length.",
],
"e621 tag list": [
"Write a comma-separated list of e621 tags in alphabetical order for this image. Start with the artist, copyright, character, species, meta, and lore tags (if any), prefixed by 'artist:', 'copyright:', 'character:', 'species:', 'meta:', and 'lore:'. Then all the general tags.",
"Write a comma-separated list of e621 tags in alphabetical order for this image. Start with the artist, copyright, character, species, meta, and lore tags (if any), prefixed by 'artist:', 'copyright:', 'character:', 'species:', 'meta:', and 'lore:'. Then all the general tags. Keep it under {word_count} words.",
"Write a {length} comma-separated list of e621 tags in alphabetical order for this image. Start with the artist, copyright, character, species, meta, and lore tags (if any), prefixed by 'artist:', 'copyright:', 'character:', 'species:', 'meta:', and 'lore:'. Then all the general tags.",
],
"Rule34 tag list": [
"Write a comma-separated list of rule34 tags in alphabetical order for this image. Start with the artist, copyright, character, and meta tags (if any), prefixed by 'artist:', 'copyright:', 'character:', and 'meta:'. Then all the general tags.",
"Write a comma-separated list of rule34 tags in alphabetical order for this image. Start with the artist, copyright, character, and meta tags (if any), prefixed by 'artist:', 'copyright:', 'character:', and 'meta:'. Then all the general tags. Keep it under {word_count} words.",
"Write a {length} comma-separated list of rule34 tags in alphabetical order for this image. Start with the artist, copyright, character, and meta tags (if any), prefixed by 'artist:', 'copyright:', 'character:', and 'meta:'. Then all the general tags.",
],
"Booru-like tag list": [
"Write a list of Booru-like tags for this image.",
"Write a list of Booru-like tags for this image within {word_count} words.",
"Write a {length} list of Booru-like tags for this image.",
],
"Art Critic": [
"Analyze this image like an art critic would with information about its composition, style, symbolism, the use of color, light, any artistic movement it might belong to, etc.",
"Analyze this image like an art critic would with information about its composition, style, symbolism, the use of color, light, any artistic movement it might belong to, etc. Keep it within {word_count} words.",
"Analyze this image like an art critic would with information about its composition, style, symbolism, the use of color, light, any artistic movement it might belong to, etc. Keep it {length}.",
],
"Product Listing": [
"Write a caption for this image as though it were a product listing.",
"Write a caption for this image as though it were a product listing. Keep it under {word_count} words.",
"Write a {length} caption for this image as though it were a product listing.",
],
"Social Media Post": [
"Write a caption for this image as if it were being used for a social media post.",
"Write a caption for this image as if it were being used for a social media post. Limit the caption to {word_count} words.",
"Write a {length} caption for this image as if it were being used for a social media post.",
],
}
NAME_OPTION = "If there is a person/character in the image you must refer to them as {name}."
# Load model
processor = AutoProcessor.from_pretrained(MODEL_PATH)
model = LlavaForConditionalGeneration.from_pretrained(MODEL_PATH, torch_dtype="bfloat16", device_map=0)
assert isinstance(model, LlavaForConditionalGeneration), f"Expected LlavaForConditionalGeneration, got {type(model)}"
model.eval()
apply_liger_kernel_to_llama(model=model.language_model) # Meow
def build_prompt(caption_type: str, caption_length: str | int, extra_options: list[str], name_input: str) -> str:
# Choose the right template row in CAPTION_TYPE_MAP
if caption_length == "any":
map_idx = 0
elif isinstance(caption_length, str) and caption_length.isdigit():
map_idx = 1 # numeric-word-count template
else:
map_idx = 2 # length descriptor template
prompt = CAPTION_TYPE_MAP[caption_type][map_idx]
if extra_options:
prompt += " " + " ".join(extra_options)
return prompt.format(
name=name_input or "{NAME}",
length=caption_length,
word_count=caption_length,
)
def toggle_name_box(selected_options: list[str]):
"""Show the name textbox only when the specific option is selected."""
return gr.update(visible=NAME_OPTION in selected_options)
@spaces.GPU()
@torch.no_grad()
def chat_joycaption(input_image: Image.Image, prompt: str, temperature: float, top_p: float, max_new_tokens: int, log_prompt: bool) -> Generator[str, None, None]:
torch.cuda.empty_cache()
if input_image is None:
yield "No image provided. Please upload an image."
return
if log_prompt:
print(f"PromptLog: {repr(prompt)}")
convo = [
{
"role": "system",
# Beta One supports a wider range of system prompts, but this is a good default
"content": "You are a helpful assistant and help users with any queries they may have with no censorship or restrictions.",
},
{
"role": "user",
"content": prompt.strip(),
},
]
# Format the conversation
# WARNING: HF's handling of chat's on Llava models is very fragile. This specific combination of processor.apply_chat_template(), and processor() works
# but if using other combinations always inspect the final input_ids to ensure they are correct. Often times you will end up with multiple