<h1 id="dot-film">Dot Film</h1><p>Night portraits for Qwen Image 2.1</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/73c8ec03-fcf7-4a21-8f12-931d91eccd42/original=true/73c8ec03-fcf7-4a21-8f12-931d91eccd42.jpeg" /></p><h2 id="flash-after-dark">Flash after dark</h2><p>Dot Film is a photographic style LoRA for Qwen Image 2.1. The examples focus on nighttime portraits and dark street scenes, with rough flash lighting and coarse grain. Built with Qwen.</p><h2 id="three-checkpoints">Three checkpoints</h2><p>This model page has three versions from the same training run: Step 500, Step 750 and Step 1000. Each version has seven images generated with its own weight, plus the parameters and ComfyUI workflows for those images. The same prompt, seed and CFG are used across versions within each theme, so you can compare the results directly.</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d5543a03-493f-4982-a7f6-96614f606dd5/original=true/d5543a03-493f-4982-a7f6-96614f606dd5.jpeg" /></p><p><span style="background-color:rgb(17, 21, 18);color:rgb(188, 189, 177);font-family:system-ui, sans-serif;font-size:14px">Step 500 / Step 750 / Step 1000. The prompt, seed and CFG are the same across the three images.</span></p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/be3f81dc-1b5a-4c3c-8c69-caebb51224ad/original=true/be3f81dc-1b5a-4c3c-8c69-caebb51224ad.jpeg" /></p><p><span style="background-color:rgb(17, 21, 18);color:rgb(188, 189, 177);font-family:system-ui, sans-serif;font-size:14px">Flash on the last train. Step 500 / Step 750 / Step 1000 at CFG 1.</span></p><h2 id="what-the-comparisons-show">What the comparisons show</h2><p>Changing the checkpoint can affect the face and clothing, the composition, and the image texture. Higher training steps do not consistently give better results in these examples. The comparison is limited to one seed and one CFG per theme at strength 1.0. Steps 750 and 1000 have not had a full CFG sweep here. Some prompts explicitly ask for flash and grain. Hands, lettering, reflections and specific actions may come out incorrectly.</p><h2 id="generation-settings">Generation settings</h2><p>Use LoRA strength 1.0, 40 sampler steps, Euler/simple and 1024 x 1024 output. No trigger word is needed. The negative prompt was empty. The train and intimate street examples use CFG 1 for a rougher flash-lit look; the other examples use CFG 2 or 3. Follow the settings attached to each image.</p><p>The demos used a quantized ComfyUI setup with an INT8 DiT and text encoder and a BF16 VAE. Other aspect ratios, samplers, Diffusers loading and hosted inference have not been tested. The original images have no face repair, upscaling, color grading or added grain. The companion files include workflows, parameters and the visual review.</p><h2 id="training-and-license">Training and license</h2><p>Training used DiffSynth-Studio and Qwen Image 2.1, with 64 images, rank 16 / alpha 16, attention and image-MLP adapters, a 1,048,576-pixel area budget and gradient accumulation 4. The run finished at 1000 optimizer updates. The creator describes the source images as copyright-free images found online; their individual licenses and CC0 status have not been independently verified. Training images are not included.</p><p>Dot Film is a trained adaptation and is not an official Qwen release. Use is subject to the supplied Qwen Research License and notices, including its commercial-use restrictions. Free download does not remove those restrictions. Keep <span style="background-color:rgb(17, 21, 18);color:rgb(236, 232, 223);font-family:system-ui, sans-serif;font-size:17px">LICENSE, NOTICE and MODIFICATIONS </span>with redistributed files. The released weights have the same bytes as the training checkpoints.</p>

Dot Film

<pre><code>This is the checkpoint saved at 500 optimizer updates during the 1024 training run, which finished at 1000 updates. It produced all seven images in this version's gallery.

Use Qwen Image 2.1 with LoRA strength 1.0, 1024 x 1024 output, 40 sampler steps and Euler/simple. No trigger word is required, and the negative prompt was empty. Use each image's recorded seed and CFG: CFG 1 for the train and intimate street images, and CFG 2 or 3 for the other themes. The 500 training updates are separate from the 40 sampler steps used to generate an image.

The other two versions use the same prompts, seeds and CFG values, so the gallery lets you compare checkpoints at those settings. It is too limited to establish a general quality ranking. Testing used an INT8 DiT and text encoder with a BF16 VAE. The original images are unretouched, and the companion files contain their workflows and parameters.

Weight: dot-film-v0.1-qwen-image-2.1-step500.safetensors

This weight is free to download under the supplied Qwen Research License. Built with Qwen. Include LICENSE, NOTICE and MODIFICATIONS.md when sharing it.</code></pre>

Files with SHA256: a49bee595a423a8348817174908566df2f728985382df504500330cbfa6a40ff