Atelier Qwen
A local image pipeline: ten thumbnails and a car shoot, no studio, no API
- thumbnails delivered for 10 real channel videos
- 10
- of total rendering for all ten, on a laptop
- 40 min 59 s
- per 1024² image, measured on our machine
- 241 s
- no API, no credits, no online service
- €0
- The problem
- The daily shorts only had a frame grabbed from the video as a thumbnail, and online generators cost a subscription and send everything out.
- For whom
- The LaRefVlog channel first; then any team wanting visuals without a studio, a stock library or an online service.
- The outcome
- Ten thumbnails delivered for ten real videos and four car advertising visuals, rendered on a 16 GB laptop, every image traced.
Context
The channel’s daily shorts only had a frame grabbed from the video as a thumbnail. They needed real thumbnails, in the channel’s format and in native vertical, without subscribing to an online generator.
The technical bet: run Qwen-Image-2.1 (7 billion parameters) on a 16 GB MacBook, measure honestly what it costs in time and memory, then use it to deliver, with production rigour.
What I built
- Two paths installed and compared: a 4-bit MLX engine (fastest) and ComfyUI with GGUF weights (most integrable, LoRAs and nodes), with memory guards calibrated on measurements.
- A six-configuration speed bench (resolution, step count), same prompt and seed, every render logged.
- Ten thumbnails for the videos of 10 to 19 October, in 16:9 matching the channel’s exact format and in 1080×1920 vertical; text is composed in post-production, never asked of the model.
- A four-visual photorealistic car shoot (studio, urban night, golden hour, concrete), with a photographic finish (grain, vignetting) and the model’s raw output archived alongside.
- A manifest per image (prompt, seed, timings, SHA-256 hashes) and a standalone presentation, ready to show a client.
- The same pipeline as a ComfyUI workflow (GGUF and 4-step Lightning LoRA): 14 nodes, titling and 16:9 format by two custom nodes; carried into the P07 node pack and run end to end on 20 Oct in 188 s, model loading included.
Architecture 6 blocks
- 01 · InputDay’s videoTitle, idea, extracted frame
- 02 · ToolRecipeShot-description prompt, seed, format
- 03 · ModelQwen-Image-2.17 billion parameters, 4-bit, MLX, 6 steps
- 04 · ToolPost-productionText, parody label, signature, photo finish
- 05 · HumanSelectionVisual approval before delivery
- 06 · OutputDelivery16:9, 9:16, manifest and hashes
The model’s raw output is always archived next to the final: what came from the model and what came from finishing can be proven.
Pipeline 5 steps
- BriefThe day’s video title and idea.
- RecipeA shot-description prompt (lens, light, material) and a fixed seed.
- RenderingQwen-Image-2.1 4-bit locally, 6 steps, one render at a time.
- Post-productionSlogan, “parody fiction” label and signature composed separately.
- Delivery16:9 and 9:16, JPEG under 2 MB, manifest and hashes.
Decisions the choice, the reason, the alternative ruled out
- D1Never ask the model for text; compose it in post-production.
Image models write poorly; composed text is crisp, in the channel’s typography, and editable without a new render.
Ruled outLetting the model write the slogan.
- D2Turbo 6 steps rather than the 25-step base model.
Measured: 2.7 times faster for a marginal visual gain on these cases.
- D3Describe a real shot (lens, aperture, light, material) in every recipe.
That vocabulary pushes the model towards photography rather than illustration.
- D4Memory guards calibrated on our measurements, one render at a time.
On 16 GB, one render too many tips the machine into swap; better to refuse cleanly and rerun.
Results
Speed bench (MacBook Pro M1 Pro 16 GB, same prompt, same seed)
| Configuration | Total time | Per step | Published reference (M4) |
|---|---|---|---|
| Turbo 6 steps, 512² | 94 s | 8.1 s | — |
| Turbo 6 steps, 768² | 184 s | 22.0 s | — |
| Turbo 6 steps, 1024² | 241 s | 31.6 s | 139–145 s |
| Turbo 5 steps, 1024² | 173 s | 31.2 s | — |
| Turbo 8 steps, 1024² | 255 s | 29.9 s | — |
| Base 25 steps, 1024² | 648 s | 25.3 s | 536 s |
Turbo 6 steps is the best trade-off: base 25 steps takes 2.7 times longer for a marginal visual gain. Published reference measured on an M4.
Deliverables 5
- Ten 16:9 thumbnails and ten 1080×1920 vertical masters, with before/after boards.
- Four car shoot visuals, raw and final, with finishing settings.
- A six-configuration speed bench and its memory reports.
- Client guides, reusable recipes and a standalone presentation.
- The thumbnail ComfyUI workflow (14 nodes, two custom titling nodes), run on 20 Oct through the P07 pack.
Environment
Hardware and measurement setup
- Measured on a MacBook Pro M1 Pro 16 GB: the ~7.7 GB model goes through swap; faster on an M4 (published reference: 139–145 s at 1024²).
- One render at a time, under memory guards; existing LoRAs used at inference, training a LoRA for this model belongs on a dedicated GPU.
Gallery 14
What this project means for you
The same know-how, on your problem.
Your visuals without a studio
Thumbnails, product or campaign visuals made on your machines, with your recipes and no subscription.
A benchmark before you buy
What a model really costs in time and memory on your hardware, measured, before you invest.
Full traceability
Prompt, seed, timings and hashes for every image: you know where every visual came from.