AI Desktop Girlfriend in ComfyUI – MiniMax H3 (Workflows Included)
This workflow is free for a limited time! You may have seen the “AI Desktop Girlfriend” videos on social media: A lively woman sits on a sofa inside a desktop screen, alongside the application icons and the taskbar. She looks and smiles at you, chat with you, do what you ask her to do, and…
Annons
Annons
This workflow is free for a limited time! You may have seen the “AI Desktop Girlfriend” videos on social media: A lively woman sits on a sofa inside a desktop screen, alongside the application icons and the taskbar. She looks and smiles at you, chat with you, do what you ask her to do, and sometimes throw a tantrum! Clips like this can be made with ComfyUI. Unmute for sound! In this tutorial, you will learn how to: Create the viral AI Desktop Girlfriend video on ComfyUI. Use a screenshot of your desktop background. Use any person’s or your own image. Three example scenarios: changing clothes, answering a question, and messing up your desktop icons. Software needed This guide uses ComfyUI, a free AI image and video generator that you can run locally on Windows, Mac, or online through Google Colab. Think Diffusion offers a ready-to-use online ComfyUI service if you don’t have a powerful GPU. (SDA readers get 20% extra credit.) How the AI Desktop Girlfriend Workflow Works These workflows use side-by-side reference trick with the MiniMax H3 Reference-to-Video (ref2va) model. Before the images reach the MiniMax H3 Reference to Video node, ComfyUI’s built-in Stitch Images (ImageStitch) node joins two pictures side by side into one composite sheet (): Then we spell out three rules: Left side = Scene reference: Draw the Windows desktop from the left panel. Right side = Character reference: Use the woman from the right panel. No split screen: Drop the right panel from the final video. Produce a full-screen desktop recording with the woman sits naturally on the couch. Since MiniMax H3 produces video and audio at the same time, we get a consistent video with correct lip-sync and waveform. AI Desktop Girlfriend Step-by-Step Step 0: Update ComfyUI MiniMax H3 needs a recent ComfyUI build (0.30.0 or newer, and 0.36.0+ is best for fast VAE decoding). Update your installation first through ComfyUI Manager. Click the Manager button on the top toolbar. Click Update ComfyUI. Restart ComfyUI and refresh your browser tab. Step 1: Download the models Grab these five files from the official MiniMax H3 repository and drop each one into its matching folder: Download minimax_h3_ref2va_pruned_int8_convrot.safetensors (19.5 GB) and place it in ComfyUI > models > diffusion_models. Download qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (14.6 GB) and place it in ComfyUI > models > text_encoders. Download minimax_h3_video_vae_int8_convrot.safetensors (2.62 GB) and place it in ComfyUI > models > vae. Download minimax_h3_audio_vae_fp32.safetensors (577 MB) and place it in ComfyUI > models > vae. Download minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (1.82 GB) and place it in ComfyUI > models > loras. You should have these files installed if you have followed my tutorial MiniMax H3 Reference-to-Video. If minimax_h3_video_vae_fp16.safetensors is present instead of the int8_convrot video VAE, simply select that file in the top Load VAE node. Step 2: Load the workflow Pick one of the three workflow JSON files below and drag it onto your ComfyUI canvas: Change cloth: Download Answer a question: Download Mess up your icons: Download Step 3: Install missing nodes These workflows uses ComfyUI core nodes, an updated ComfyUI should load them alright. But if I am wrong and any node shows a red border, fix it with ComfyUI Manager: Click Manager > Install Missing Custom Nodes. Install whatever packages the list reports as missing. Restart ComfyUI. Step 4: Upload your desktop screenshot and character portrait Find the User Inputs group in the bottom-left corner of the graph. Two Load Image nodes there feed the Stitch Images (Desktop + Character) node: In Load Image (Desktop Scene) on the left, upload a 16:9 screenshot of your computer desktop. Use a similar dark theme and a couch for the best result with the preset prompt. In Load Image (Character Reference) on the right, upload a clear half-body character. You can also use a photo of yourself… haha. Feel free to use mine for testing: Input 1: Desktop background Input 2: The character With direction: right and match_image_size: true, the Stitch Images (Desktop + Character) node joins both pictures side by side and sends the combined sheet to ref_image_0 () on MiniMax H3 Reference to Video. Step 5: Set the resolution and duration In the User Inputs group, verify the video size and length: On Resolution Selector (Size), leave aspect_ratio at 16:9 (Widescreen). Set megapixels to 0.4 (864x480) when you experiment with prompts and seeds, or raise it to 0.98 (1344x768) for a full 768p video. Set multiple at 32. Set Float (Duration) to 8 seconds for the outfit-change workflow (192 frames at 24fps), or 6 seconds for the question-and-answer and desktop-icon workflows (141 frames at 24fps). Step 6: Customize the prompt Every workflow JSON ships with the full prompt pre-loaded in the Input Text (Prompt) node. If you would rather write your own scenario, keep the prompt organized into these five sections: Reference Image Lock: contains two regions (LEFT SIDE = FINAL SCENE REFERENCE and RIGHT SIDE = CHARACTER REFERENCE). Make it explicit that the final video is not split-screen and must restore the full-screen 16:9 desktop with inside it. Style & Static Camera: Define a locked static camera (single continuous shot, no cuts, no zoom, no pan) and require the taskbar and icon untouched. (adjust depending on your video) Subtitle & Waveform UI: Tell the model to put the floating voice overlays in the LEFT-CENTER above the couch: light-purple text with a speaker icon when the woman ((S1)) speaks, white text with a microphone icon when the off-screen male voice ((S2)) speaks, and a waveform below. Timecoded Action & Dialogue ( tags): Divide the timeline into 2-to-3-second beats (At 00:00.000, At 00:02.200, At 00:05.200). Wrap spoken lines in [English] ... tags so the speech and lip-sync stay aligned. Soundscape (overall_soundscape): List the physical cues at the bottom, such as sofa movement, footsteps walking off-screen, fabric rustle, or mouse clicks. Set non_diegetic_music: N/A so background music never drowns out the speech. Step 7: Generate the video Press the Run button (or Ctrl + Enter) to start rendering. Once the job completes, the video plays right inside the Save Video node on the right and is saved to ComfyUI > output > video. Settings SettingRecommended ValueNotes & EffectsStitch Directionright (match_image_size: true)Join the two images left (desktop) and right (character)ref_image_sizematchResize the reference image to match the target size (max preserves up to a 2048px short edge but runs 3x slower).Aspect Ratio16:9 (Widescreen)Generated video’s aspect ratio.Megapixels0.4 (864x480) or 0.98 (1344x768)0.4 MP for drafts and 0.98 MP for high quality finals.Multiple32Don’t change this.Duration6 to 8 secondsLength of the video.Enable Lightning LoRAfalse (20 steps)Set it to true to use the 4-step Turbo LoRA for quick video generation with lower quality.Sampler / Schedulerres_multistep / betaThe beta scheduler keeps facial features slightly better than simple.Denoise1.0Don’t change this. Tips Two-reference trick The MiniMax H3 Reference to Video (ref2va) node treats the stitched image as a reference sheet. Define the two reference images in the prompt. Follow the off-screen “black-box” rule for outfit changes: Don’t ask the model to change the character’s clothes while she sits still. It won’t go well… Always have her walk completely out of the screen and come back with a new outfit. Keep her lips closed during off-screen male questions: While the off-screen male ((S2)) speaks, say explicitly in the prompt that keeps her lips closed, smiles gently, and listens. Without this instruction, the model will sometimes lip sync the male’s voice. Keep the left wall dark and uncluttered: Place the character on the right half of the leather sofa and leave the upper-left corner dark. This gives MiniMax H3 a high-contrast canvas for crisp top-left desktop icons and the floating voice waveform without covering her face. Avoid the 1.0 Megapixels setting at 16:9: Native 768p for MiniMax H3 is 1344x768, which equals 0.98 megapixels. Choosing 1.0 megapixels stretches the canvas to 1376x768, past the training resolution , and softens the facial details. Use an AI upscaler for high resolution. What’s next MiniMax H3 Reference-to-Video MiniMax H3 Image-to-Video MiniMax H3 Text-to-Video MiniMax H3 Fun ControlNet Union Kandinsky 6.0 Video with Audio Wan 2.2 Image-to-Video The post AI Desktop Girlfriend in ComfyUI – MiniMax H3 (Workflows Included) first appeared on Stable Diffusion Art.