bytedance/bagel

🥯ByteDance Seed's Bagel Unified multimodal AI that generates images, edits images, and understands images in one 7B parameter model🥯

Capabilities

Reference Images Seed

Community model (estimated from hardware time)

Name	Type	Description	Default	Constraints
`prompt` *	string	Text prompt for generation, editing, or understanding	`—`	—
`cfg_img_scale`	number	Image guidance scale for preserving input image details	`1.5`	min: 1, max: 10
`cfg_renorm_min`	number	Minimum CFG renorm value	`1`	min: 0, max: 1
`cfg_renorm_type`	string	CFG renormalization method	`"global"`	global local text_channel
`cfg_text_scale`	number	Text guidance scale for how closely to follow the prompt	`4`	min: 1, max: 20
`enable_thinking`	boolean	Enable chain-of-thought reasoning for better results	`false`	—
`image`	string (uri)	Input image for editing or understanding tasks	`—`	—
`num_inference_steps`	integer	Number of denoising steps	`50`	min: 1, max: 100
`output_format`	string	Output image format	`"webp"`	webp jpg png
`output_quality`	integer	Image compression quality for lossy formats	`90`	min: 1, max: 100
`seed`	integer	Random seed for reproducible results	`—`	—
`task`	string	Task to perform	`"text-to-image"`	text-to-image image-editing image-understanding
`timestep_shift`	number	Distribution of denoising steps between composition and details	`3`	min: 1, max: 10

prompt required string

Text prompt for generation, editing, or understanding

cfg_img_scale number

Image guidance scale for preserving input image details

Default: 1.5 min: 1, max: 10

cfg_renorm_min number

Minimum CFG renorm value

Default: 1 min: 0, max: 1

cfg_renorm_type string

CFG renormalization method

Default: "global"

global local text_channel

cfg_text_scale number

Text guidance scale for how closely to follow the prompt

Default: 4 min: 1, max: 20

enable_thinking boolean

Enable chain-of-thought reasoning for better results

Default: false

image string

Input image for editing or understanding tasks

num_inference_steps integer

Number of denoising steps

Default: 50 min: 1, max: 100

output_format string

Output image format

Default: "webp"

webp jpg png

output_quality integer

Image compression quality for lossy formats

Default: 90 min: 1, max: 100

seed integer

Random seed for reproducible results

task string

Task to perform

Default: "text-to-image"

text-to-image image-editing image-understanding

timestep_shift number

Distribution of denoising steps between composition and details

Default: 3 min: 1, max: 10

Version: 7dd8def79e50 Updated: 2/26/2026 272.8K runs