Media lab
Choose the layer before choosing the model.
A model is the learned generator. A UI is how you operate it. A runtime is what loads it. A host is where the computer lives. Keeping those layers separate makes experimentation much less confusing.
Paths by ambition
| Path | Good first choice | Strength | Tradeoff |
|---|---|---|---|
| Easy local image | Fooocus + SDXL | Offline, simple prompt-focused workflow, modest setup. | Fooocus is now limited long-term support with bug fixes; its model family is SDXL, not every new image model. |
| Controlled local pipeline | ComfyUI | Nodes expose model, conditioning, control, upscaling, video, audio, and other stages. | More power means more concepts, workflow files, custom nodes, and maintenance. |
| Hosted image/video/audio | Provider API or managed workspace | Little local setup, access to large models, easy burst capacity. | Ongoing cost, provider terms, data transfer, rate limits, and less control. |
| Cloud GPU workstation | Rented GPU Pod or dedicated GPU server | More VRAM and batch throughput without buying hardware. | Storage, security, idle billing, model downloads, and shutdown discipline matter. |
Model families worth understanding
| Family / example | Media | Practical read | License note |
|---|---|---|---|
| SDXL 1.0 | Image | Mature base model and a natural match for Fooocus. The model card describes base and optional refiner stages. | CreativeML Open RAIL++-M; read the actual model card and restrictions before redistribution or a commercial service. |
| FLUX.2 Klein | Image and editing | Newer open-weight direction; the official overview lists a 4B version under Apache 2.0 and a 9B version under a non-commercial license. | “Open weights” does not mean every size or use has the same license. Verify the exact checkpoint. |
| FLUX [dev] | Image | Strong image-generation direction with self-hosting options. | Terms can require commercial licensing, filtering, or review. Do not choose it for publishing until the exact terms are checked. |
| Video, audio, and 3D ecosystems | Video, speech/music, 3D | Fast-moving families are often easiest to operate through ComfyUI workflows or hosted endpoints. | Model cards, datasets, voices, likenesses, and outputs can have separate restrictions. Review each component. |
A saved model state
Usually a large file containing learned weights. Think of it as the trained “brain” a runtime loads. Different checkpoints can change style, anatomy, speed, memory needs, and license.
A small adaptation
A lightweight add-on that nudges a base model toward a style, subject, or behavior. It is not automatically compatible with every checkpoint, and its own license matters too.
A model architecture
Stable Diffusion XL is a diffusion-based image model family. Fooocus is an application built around SDXL; SDXL is not the same thing as Fooocus.
Graphics memory
The working space for model weights and intermediate images. Quantization, offloading, lower resolution, fewer stages, and smaller models can reduce pressure, but each changes speed or quality.
Copyright and safety checklist
- Use official model pages or trustworthy model cards, not random download mirrors.
- Record the exact model, version, checkpoint hash when available, LoRAs, prompt, seed, and source references for reproducibility.
- Read model, LoRA, dataset, voice, and tool licenses separately.
- Do not assume “local” means “permission-free.” Local generation changes data exposure, not copyright or legal duties.
- For public services, add input/output filtering, abuse handling, rate limits, private storage, and a human escalation path.
- Validate the delivered asset itself: composition, anatomy, transparency, style match, audio cleanliness, or video artifacts.
Recommended exploration order
1) one SDXL image locally, 2) one project-referenced image set, 3) a repeatable ComfyUI workflow, 4) compare one rented GPU session, 5) only then evaluate video/audio/3D or public serving. This order keeps the learning curve and spend visible.