I made two AI action films of myself on a MacBook M5 Max. Then I counted the mistakes.

Two 30-second action films starring me, made on my MacBook Pro (M5 Max, 128 GB) with LTX 2.3 in ComfyUI. Every frame, the music and every sound effect were generated on the laptop, with no cloud. The voice in the films and the narration is my own HeyGen voice clone. Then I go through both films frame by frame and count everything that went wrong: twelve mistakes, plus bloopers. The big cloud video models are far ahead of this. This is what you can do today, for free, on your own computer, with your own face. AI disclosure: all film footage is AI-generated. The face is mine and the voice is my own clone, used with my consent. "The M5 Ultra Heist" is a homage to a famous heist scene. No footage or music from the film. CHAPTERS 0:00 Cold open 0:47 Honest expectations and the 5-second face test 1:27 What you need, and the licence catch 2:15 Step 1: install ComfyUI and fix the black video 2:51 Step 2: the models 3:15 Step 3: keyframes that keep my face 3:56 Step 4: the LTX 2.3 shot and how to prompt it 4:49 Film 1: The Jump 5:18 Spot the mistakes: The Jump 7:04 Step 5: my own voice 8:04 Film 2: The M5 Ultra Heist 8:34 Spot the mistakes: The Heist 10:10 Bloopers 10:45 How I'd fix it next time 11:47 The numbers, and LTX 2.3 vs LTX-2.5 12:27 Wrap-up THE RECIPE (checked 12 Oct 2026; model hosts change files often, so re-check) Hardware: MacBook Pro, M5 Max, 128 GB unified memory, macOS 27.2. One 8-second shot peaked at 70 GB; with 64 GB, make shorter shots or drop the resolution. Software: ComfyUI v0.39.2 (github.com/Comfy-Org/ComfyUI), Python 3.11, PyTorch 2.13 on MPS, the ComfyUI-GGUF custom node (github.com/city96/ComfyUI-GGUF), ffmpeg, audio-separator 0.47.0. Launch: PYTORCH_ENABLE_MPS_FALLBACK=1 python main.py --reserve-vram 10 --disable-smart-memory Black-video fix on Apple Silicon (ComfyUI issue #15804, still open on 12 Oct 2026): in comfy/ldm/modules/sub_quadratic_attention.py change the four torch.empty(1, 1, 1, ...) calls to torch.zeros(1, 1, 1, ...). A ComfyUI update overwrites the patch, so check the issue first. Model files (folder inside ComfyUI/models): - ltx-2.3-22b-distilled-Q8_0.gguf (diffusion_models), unsloth/LTX-2.3-GGUF - gemma_3_12B_it.safetensors (text_encoders), Comfy-Org/ltx-2 - ltx-2.3-22b-distilled_embeddings_connectors.safetensors (checkpoints), unsloth/LTX-2.3-GGUF - ltx-2.3-22b-distilled_video_vae.safetensors (vae), unsloth/LTX-2.3-GGUF - ltx-2.3-22b-distilled_audio_vae.safetensors (checkpoints), unsloth/LTX-2.3-GGUF - ltx-2.3-spatial-upscaler-x2-1.1.safetensors (latent_upscale_models), Lightricks/LTX-2.3 Keyframes: an image-edit template in ComfyUI with my photo as the reference, sampled at 1280x704 (the video size). Write the prompt like a film still and end it with "Keep the man's face exactly as in the reference photo", spelling out face shape, hair, stubble and eye colour. 2 to 4.5 minutes each. Shots: ComfyUI's LTX 2.3 image-to-video template with the Q8 GGUF swapped in. 1280x704, 24 fps, 193 frames for 8 s. Two stages: 640x352 with 8 steps, a x2 latent upscale, then 3 refine steps. Euler, CFG 1, first-frame strength 0.7 then 1.0. 7:27 to 8:52 per 8 s shot: about a minute of compute per second of video. Save video and audio separately, then mux with ffmpeg. Prompt rules: one paragraph, present tense, 100 to 200 words. Camera first. Describe what the body does, not what the person feels. Put the line in quotes. End with "Audio: ..." and describe every sound. Voice: audio-separator splits each shot's audio into voice and background; my HeyGen clone reads the same lines, and each phrase is placed where the original started. Render time: The Jump about 44 minutes, the heist about 41 minutes for the shots used (about 100 minutes with every take). LICENCES - LTX 2.3 and LTX-2.5 (Lightricks): LTX-2 Community License, free for individuals and companies under US$10M annual revenue. Read it before you use it. - MiniMax H3 is not used or shown: its licence lists the UK and the EU among its Excluded Territories. - ComfyUI, ComfyUI-GGUF, audio-separator and ffmpeg are open source. Next: the same films on the M5 Ultra, with the fixes. #LocalAI #AIVideo #LTX23