Checkpoint vs LoRA vs embedding vs ControlNet: which file does what

Q: Downloaded a "LoRA" and it is 6 GB

It is a checkpoint; the uploader labelled it loosely. Put it in the checkpoints folder and select it as the model. The catalog label here is reliable.

Q: Colours look washed out

Missing or wrong VAE. Load the family's VAE; on SDXL the fp16-fix VAE.

Konu: Modeller

Quartermaster tarafından

2025-10-08 tarihinde yayımlandı

The six kinds of model files you will meet, what each one changes, how big it is, where it goes and when to use it: checkpoints and fine-tunes, LoRAs and their LyCORIS cousins, textual inversion embeddings, ControlNets and IP-Adapters, VAEs and upscalers.

Bu sayfa henüz çevrilmedi, bu yüzden İngilizce gösteriliyor.

Genel bakış

A new user downloads a "model" and discovers it is one of six different things that go in six different folders and do six different jobs. The catalog on this site labels every entry with its type (base model,

LoRA

, embedding, other) and its family for that reason. This guide explains the types once, so the labels make sense.

The rule of thumb: a checkpoint is the whole painter; a LoRA teaches the painter a new subject or style; an embedding teaches the painter a new word; a

ControlNet

or IP-Adapter tells the painter where things go or what to copy from a reference; the VAE is the painter's eyes converting between the canvas and the compressed working space; an

upscaler

is a separate craftsman who enlarges the finished picture. Only the checkpoint can work alone. Everything else needs a checkpoint of the matching family.

Sizes tell you a lot. 2 to 7 GB for an

SDXL

checkpoint and 12 to 24 GB for

Flux

; 10 to 700 MB for a LoRA; under 1 MB for an embedding; 1 to 2.5 GB for a ControlNet; 300 MB for a VAE; 60 MB for an ESRGAN upscaler. If a file called a LoRA is 6 GB, it is a checkpoint. The model page here shows the size and the hash so you can tell before downloading.

Başvuru

Ad	Tür	Ne olduğu
Checkpoint (base model)	2-24 GB	The full denoiser, usually with text encoder and VAE bundled. Fine-tunes (DreamShaper, Juggernaut, Illustrious , Pony) are checkpoints retrained on new data. Folder: models/checkpoints ( ComfyUI ), models/Stable-diffusion (A1111, Forge ). Flux, Qwen and Wan ship as separate diffusion_models + text_encoders + vae files.
LoRA	10-700 MB	A small weight difference applied on top of a checkpoint. Characters, styles, concepts, sliders, motion. Needs a trigger word and a weight. Folder: models/loras or models/Lora.
LyCORIS / LoCon / LoHa / DoRA	10-700 MB	LoRA variants that touch more layers or use a different factorisation. Loaded like LoRAs in every current UI.
Embedding (textual inversion)	10-200 KB	A learned token for the text encoder. Called by its file name in the prompt. Common as negative embeddings (bad-hands, easynegative) on SD 1.5 and SDXL; rare on Flux. Folder: models/embeddings.
ControlNet / T2I-Adapter	0.3-2.5 GB	A side network that conditions the denoiser on a pose, depth map, edge map or scribble. Family specific. Folder: models/controlnet.
IP-Adapter / reference	50 MB-1 GB	Conditions on an image instead of text: style transfer, face reference. Needs a matching CLIP vision model.
VAE	160-330 MB	Encoder/decoder between pixels and latents. Usually bundled in checkpoints; separate files fix washed-out colours (SDXL fp16-fix) or come with new families (Flux ae.safetensors, Wan VAE).
Upscaler	2-70 MB	ESRGAN, Real-ESRGAN, SwinIR, DAT: separate networks that enlarge pixels after generation. Folder: models/upscale_models.

No folders, no file types, no VAE hunting

BitVector Prism preloads the checkpoint, the LoRAs, the VAE and the ControlNets that belong together. Pick a model from this site, pick an add-on and generate in the browser.

BitVector’ı dene

Adım adım

Identify the file type from the catalog label and the size before downloading. Base model, LoRA, embedding or other on the card; the exact bytes on the model page.
Match the family. Every non-checkpoint file names the family it was built for (SDXL, Flux,
Wan 2.2
14B); it only works with checkpoints of that family.
Put the file in the folder for its type (table above) and refresh the UI.
Checkpoint: select it as the model. For split families (Flux, Qwen, Wan,
Z-Image
) load the diffusion model, the text encoders and the VAE separately in ComfyUI.
LoRA: Load LoRA node after the checkpoint, or <lora:name:0.8> in the prompt, plus the trigger word.
Embedding: type the file name in the prompt (or the
negative prompt
for negative embeddings); ComfyUI uses embedding:name.
ControlNet: Apply ControlNet node with the preprocessed image (pose, depth); start with strength 0.8 and end percent 0.8.
VAE and upscaler: select the VAE in the loader or settings when colours are wrong; run the upscaler as a final step through Upscale Image (using Model).

Örnekler

A full SDXL stack in a Forge prompt

Checkpoint: juggernautXL_v9.safetensors | VAE: sdxl_vae_fp16fix.safetensors

Prompt: zavy cinematic, <lora:zavy-cinematic-xl:0.7> a knight resting by a campfire, night, embers, 35mm

Negative: easynegative, bad-hands-5, blurry | ControlNet: openpose, weight 0.8 | Upscaler: 4x-UltraSharp 2x

The same stack as ComfyUI nodes

Load Checkpoint -> Load LoRA -> Apply ControlNet (OpenPose image) -> KSampler -> VAE Decode (fp16-fix VAE) -> Upscale Image (using Model: 4x-UltraSharp)

İpuçları

The catalog type "other" covers ControlNets, VAEs,
upscalers
and workflows; the model page says which.
Merges are checkpoints made by averaging other checkpoints; they inherit the family of their parents and keep LoRA compatibility.
An
inpainting
checkpoint is a checkpoint with extra input channels; it needs the inpainting workflow and is not a drop-in for text-to-image.
Negative embeddings were the SD 1.5 way of fixing hands; on SDXL they help less and on Flux they do nothing because the model has no
CFG
negative.
Flux Fill, Flux Redux, Flux Depth and Canny are checkpoints or adapters, not LoRAs, even though they ship beside the LoRAs on the hub.
When in doubt, the cloud partners have the right combination preloaded: pick the model and the LoRA and the matching VAE and ControlNet are handled for you.

Use the fp16-fix VAE or the no-half-vae flag; on Flux check that the text encoder precision matches the build.

Checkpoints, LoRAs, embeddings and ControlNets, already wired

PirateDiffusion keeps thousands of model files organised for you on its own GPUs; you name the checkpoint and the LoRA tags in a Telegram command and the matching VAE and ControlNets are applied. Unlimited generation, fixed price.

PirateDiffusion’ı dene

Sorular

Which one should a beginner download first?

One checkpoint of the family you want to use. LoRAs come second, once you know which base you like. Embeddings and ControlNets come when a specific problem needs them.