lucataco/thinkdiffusionxl – Run with an API on Replicate

lucataco / thinkdiffusionxl

ThinkDiffusionXL is a go-to model capable of amazing photorealism that's also versatile enough to generate high-quality images across a variety of styles and subjects without needing to be a prompting genius

Cold

Public
15.8K runs
L40S
GitHub
Paper
License

Iterate in the playground

Run with an API

Playground API Examples README Versions

Input

prompt

string

Shift + Return to add a new line

cinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeouscinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeous

Input prompt

Default: "An astronaut riding a rainbow unicorn"

negative_prompt

string

Shift + Return to add a new line

anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokehanime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokeh

Input Negative Prompt

Default: ""

width

integer

Width of output image

Default: 1024

height

integer

Height of output image

Default: 1024

num_outputs

integer

(minimum: 1, maximum: 4)

Number of images to output.

Default: 1

scheduler

string

scheduler

Default: "DPMSolverMultistep"

num_inference_steps

integer

(minimum: 1, maximum: 500)

Number of denoising steps

Default: 30

guidance_scale

number

(minimum: 1, maximum: 50)

Scale for classifier-free guidance

Default: 7

seed

integer

Random seed. Leave blank to randomize the seed

nsfw_checker

boolean

Select to filter for NSFW content

Default: true

Run this model in Node.js with one line of code:

npx create-replicate --model=lucataco/thinkdiffusionxl

or set up a project from scratch

Install Replicate’s Node.js client library:

npm install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import and set up the client:

import Replicate from "replicate";

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,
});

Run lucataco/thinkdiffusionxl using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

const output = await replicate.run(
  "lucataco/thinkdiffusionxl:c41c12756b561bc81047a9307c9143589d158ef084655dbb3073b4f08ff54f32",
  {
    input: {
      seed: 403,
      width: 1024,
      height: 1024,
      prompt: "cinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeous",
      scheduler: "DPMSolverMultistep",
      num_outputs: 1,
      nsfw_checker: true,
      guidance_scale: 6,
      negative_prompt: "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokeh",
      num_inference_steps: 30
    }
  }
);

// To access the file URL:
console.log(output[0].url()); //=> "http://example.com"

// To write the file to disk:
fs.writeFile("my-image.png", output[0]);

To learn more, take a look at the guide on getting started with Node.js.

Install Replicate’s Python client library:

pip install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import the client:

import replicate

Run lucataco/thinkdiffusionxl using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

output = replicate.run(
    "lucataco/thinkdiffusionxl:c41c12756b561bc81047a9307c9143589d158ef084655dbb3073b4f08ff54f32",
    input={
        "seed": 403,
        "width": 1024,
        "height": 1024,
        "prompt": "cinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeous",
        "scheduler": "DPMSolverMultistep",
        "num_outputs": 1,
        "nsfw_checker": True,
        "guidance_scale": 6,
        "negative_prompt": "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokeh",
        "num_inference_steps": 30
    }
)
print(output)

To learn more, take a look at the guide on getting started with Python.

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Run lucataco/thinkdiffusionxl using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

curl -s -X POST \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait" \
  -d $'{
    "version": "lucataco/thinkdiffusionxl:c41c12756b561bc81047a9307c9143589d158ef084655dbb3073b4f08ff54f32",
    "input": {
      "seed": 403,
      "width": 1024,
      "height": 1024,
      "prompt": "cinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeous",
      "scheduler": "DPMSolverMultistep",
      "num_outputs": 1,
      "nsfw_checker": true,
      "guidance_scale": 6,
      "negative_prompt": "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokeh",
      "num_inference_steps": 30
    }
  }' \
  https://api.replicate.com/v1/predictions

To learn more, take a look at Replicate’s HTTP API reference docs.

Output

{
  "completed_at": "2023-11-06T23:25:54.366426Z",
  "created_at": "2023-11-06T23:25:46.334249Z",
  "data_removed": false,
  "error": null,
  "id": "fezk7clbhkwinigd5exfp7mr2q",
  "input": {
    "seed": 403,
    "width": 1024,
    "height": 1024,
    "prompt": "cinematic film still dramatic side lighting, dramatic intense stare closeup portrait, dark black background, hdr, dramatic beautiful warrior woman with warrior face paintings and blood, viking braids, blue eyes, pelt, skull necklace, shallow depth of field, vignette, highly detailed, high budget Hollywood film, cinemascope, moody, epic, gorgeous",
    "scheduler": "DPMSolverMultistep",
    "num_outputs": 1,
    "nsfw_checker": true,
    "guidance_scale": 6,
    "negative_prompt": "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, blur, bokeh",
    "num_inference_steps": 30
  },
  "logs": "Using seed: 403\n  0%|          | 0/30 [00:00<?, ?it/s]\n  3%|▎         | 1/30 [00:00<00:05,  4.96it/s]\n 10%|█         | 3/30 [00:00<00:04,  6.18it/s]\n 13%|█▎        | 4/30 [00:00<00:04,  5.70it/s]\n 17%|█▋        | 5/30 [00:00<00:04,  5.44it/s]\n 20%|██        | 6/30 [00:01<00:04,  5.25it/s]\n 23%|██▎       | 7/30 [00:01<00:04,  5.15it/s]\n 27%|██▋       | 8/30 [00:01<00:04,  5.08it/s]\n 30%|███       | 9/30 [00:01<00:04,  5.04it/s]\n 33%|███▎      | 10/30 [00:01<00:03,  5.01it/s]\n 37%|███▋      | 11/30 [00:02<00:03,  4.98it/s]\n 40%|████      | 12/30 [00:02<00:03,  4.96it/s]\n 43%|████▎     | 13/30 [00:02<00:03,  4.95it/s]\n 47%|████▋     | 14/30 [00:02<00:03,  4.95it/s]\n 50%|█████     | 15/30 [00:02<00:03,  4.94it/s]\n 53%|█████▎    | 16/30 [00:03<00:02,  4.93it/s]\n 57%|█████▋    | 17/30 [00:03<00:02,  4.93it/s]\n 60%|██████    | 18/30 [00:03<00:02,  4.92it/s]\n 63%|██████▎   | 19/30 [00:03<00:02,  4.92it/s]\n 67%|██████▋   | 20/30 [00:03<00:02,  4.92it/s]\n 70%|███████   | 21/30 [00:04<00:01,  4.92it/s]\n 73%|███████▎  | 22/30 [00:04<00:01,  4.91it/s]\n 77%|███████▋  | 23/30 [00:04<00:01,  4.91it/s]\n 80%|████████  | 24/30 [00:04<00:01,  4.92it/s]\n 83%|████████▎ | 25/30 [00:04<00:01,  4.92it/s]\n 87%|████████▋ | 26/30 [00:05<00:00,  4.92it/s]\n 90%|█████████ | 27/30 [00:05<00:00,  4.91it/s]\n 93%|█████████▎| 28/30 [00:05<00:00,  4.91it/s]\n 97%|█████████▋| 29/30 [00:05<00:00,  4.91it/s]\n100%|██████████| 30/30 [00:05<00:00,  4.91it/s]\n100%|██████████| 30/30 [00:05<00:00,  5.02it/s]",
  "metrics": {
    "predict_time": 8.061587,
    "total_time": 8.032177
  },
  "output": [
    "https://replicate.delivery/pbxt/mfzcQjPBtjUCMaIKko5AqlwS6DL4VRB8DqBUq1fc8q8Bgy1RA/out-0.png"
  ],
  "started_at": "2023-11-06T23:25:46.304839Z",
  "status": "succeeded",
  "urls": {
    "get": "https://api.replicate.com/v1/predictions/fezk7clbhkwinigd5exfp7mr2q",
    "cancel": "https://api.replicate.com/v1/predictions/fezk7clbhkwinigd5exfp7mr2q/cancel"
  },
  "version": "c41c12756b561bc81047a9307c9143589d158ef084655dbb3073b4f08ff54f32"
}

Generated in

8.1 seconds

Tweak itReport

Examples

View more examples

Run time and cost

This model costs approximately $0.16 to run on Replicate, or 6 runs per $1, but this varies depending on your inputs. It is also open source and you can run it on your own computer with Docker.

This model runs on Nvidia L40S GPU hardware. Predictions typically complete within 3 minutes. The predict time for this model varies significantly based on the inputs.

Readme

— The following is from the ThinkDiffusion Model Card as of 11/6/23 —

ThinkDiffusionXL (TDXL) ThinkDiffusionXL is the result of our goal to build a go-to model capable of amazing photorealism that’s also versatile enough to generate high-quality images across a variety of styles and subjects without needing to be a prompting genius.

You can find it preloaded on https://www.thinkdiffusion.com

Please leave a review if you’re happy with it, this will encourage us to create more and improve on it.

The work Data source: TDXL is trained on over 10,000 diverse images that span photorealism, digital art, anime, and more. The smallest resolution in our dataset is 1365x2048, but many images go up to resolutions as high as 4622x6753. In total, our dataset takes up 42GB.

Training: With 1.8 million steps, we’ve put in the work. For comparison, Juggernaut is at 600k steps and RealVisXL is at 348k steps

Hand-captioned images: Each image is carefully captioned by hand, enhancing the model’s ability to generate accurate and high-quality results from minimal prompts.

NSFW capabilities: The model includes over 1,000 tastefully curated NSFW images.

Our thoughts Detail and quality: Most XL models in the Realistic category suffer from poor detail, especially in the background and even in basic features like eyes, teeth, and skin. We believe TDXL outperforms in these areas due to its large, high-quality dataset. For comparison, Juggernaut has about half the image material, and RealVisXL has only 1,700 images. Ultimately, TDXL simply possesses much more “knowledge”.

Less bias: We made sure to use an equal number of images for each style, gender, etc. Other models we tested over the past few months had some kind of bias, sometimes it was bias toward portrait shots, gender bias, certain ethnicities, etc. For instance, Juggernaut has a bias in the Close-Up area, and the Cinematic Light is quite dominant in that model. RealVisXL also has a bias towards Portrait shots. On the other hand, TDXL gives you what you want: Landscape, Midshot, Full Body, Close-Up, Portrait, Sideview, Backview, Action Shots, Cinematic…whatever you want without always being pushed in a certain direction due to a bias.

Versatile base: Because of its large balanced quality dataset, TDXL is versatile to serve as a base model for future trainings. You can create new finetunes in entirely different directions, add LoRAs to fill in missing concepts, or do additional trainings with more balanced quality data.