diff --git a/studio/backend/core/training/diffusion_train_common.py b/studio/backend/core/training/diffusion_train_common.py
index 5c29a6d23d..e56c3f61d5 100644
--- a/studio/backend/core/training/diffusion_train_common.py
+++ b/studio/backend/core/training/diffusion_train_common.py
@@ -295,22 +295,13 @@ _FAMILY_LABELS = {
"flux.2-dev": "FLUX.2-dev",
}
_FAMILY_VRAM_NOTES = {
- "sdxl": "Trains on ~12 GB+ (bf16 LoRA). The lightest, fastest option.",
- "flux.1": (
- "12B model, QLoRA (nf4) by default (~16 GB+). Gated on Hugging Face: accept the "
- "FLUX.1-dev license and add your HF token before training."
- ),
- "qwen-image": "20B model, QLoRA (nf4) by default (~24 GB+). The heaviest option.",
- "z-image": "6B model, QLoRA (nf4) by default (~12 GB+). bf16 only.",
- "krea-2": (
- "12B model, QLoRA (nf4) by default (~18 GB+). bf16 only. Trains on the "
- "undistilled Krea-2-Raw (Krea's guidance: train on Raw, run adapters on Turbo)."
- ),
- "flux.2-klein": "4B model, QLoRA (nf4) by default (~10 GB+). bf16 only.",
- "flux.2-dev": (
- "32B model, QLoRA (nf4) by default (~28 GB+). bf16 only. Gated on Hugging Face: "
- "accept the FLUX.2-dev license and add your HF token before training."
- ),
+ "sdxl": "~12 GB+. The lightest option.",
+ "flux.1": "12B, QLoRA by default (~16 GB+). Gated: needs its license and your HF token.",
+ "qwen-image": "20B, QLoRA by default (~24 GB+). The heaviest option.",
+ "z-image": "6B, QLoRA by default (~12 GB+).",
+ "krea-2": "12B, QLoRA by default (~18 GB+). Trains on Krea-2-Raw, runs on Turbo.",
+ "flux.2-klein": "4B, QLoRA by default (~10 GB+).",
+ "flux.2-dev": "32B, QLoRA by default (~28 GB+). Gated: needs its license and your HF token.",
}
# The flow-matching DiT families (run by diffusion_dit_trainer). They expose the base_precision /
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index f34e2f5059..5b585ae446 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -2173,11 +2173,8 @@ _DATASET_EXAMPLES: list[dict] = [
"id": "dreambooth-dog",
"label": "Dog (DreamBooth subject)",
"repo": "diffusers/dog-example",
- "description": (
- "5 photos of one dog. The classic DreamBooth subject set: teach the model a "
- "specific subject, then summon it with the trigger prompt."
- ),
- "license": "Released by Google for DreamBooth research/demos",
+ "description": "5 photos of one dog. Teach a subject, then call it with the trigger.",
+ "license": "Google, research and demos",
"image_cap": 10,
"suggested_trigger": "a photo of sks dog",
"loader": "hf_dataset",
@@ -2188,10 +2185,7 @@ _DATASET_EXAMPLES: list[dict] = [
"id": "tuxemon",
"label": "Tuxemon (captioned style set)",
"repo": "linoyts/Tuxemon",
- "description": (
- "Captioned cartoon monster art. A good style set: each image ships a caption, "
- "so the adapter learns the look without a trigger word."
- ),
+ "description": "Captioned cartoon monster art. A style set, no trigger needed.",
"license": "cc-by-sa-3.0",
"image_cap": 60,
"suggested_trigger": None,
@@ -2203,10 +2197,7 @@ _DATASET_EXAMPLES: list[dict] = [
"id": "tarot-1920",
"label": "1920 Tarot (public domain style set)",
"repo": "multimodalart/1920-raider-waite-tarot-public-domain",
- "description": (
- "Public-domain 1920 Raider-Waite tarot art with captions. A permissive style "
- "set for demoing captioned LoRA training."
- ),
+ "description": "Captioned 1920 Raider-Waite tarot art. A permissive style set.",
"license": "public domain",
"image_cap": 60,
"suggested_trigger": None,
@@ -2218,11 +2209,8 @@ _DATASET_EXAMPLES: list[dict] = [
"id": "smithsonian-butterflies",
"label": "Smithsonian Butterflies",
"repo": "huggan/smithsonian_butterflies_subset",
- "description": (
- "100 butterfly specimen photos. The classic diffusers-docs training set. No "
- "captions, so pair it with the trigger prompt to teach a butterfly subject."
- ),
- "license": "CC0 (Smithsonian Open Access)",
+ "description": "100 butterfly photos. No captions, so use the trigger prompt.",
+ "license": "CC0",
"image_cap": 100,
# The metadata columns are species names / boilerplate alt-text, not captions, so train
# it as a subject set with the trigger prompt instead.
@@ -2235,10 +2223,7 @@ _DATASET_EXAMPLES: list[dict] = [
"id": "pixel-nouns",
"label": "Nouns (pixel avatars)",
"repo": "m1guelpf/nouns",
- "description": (
- "100 captioned Nouns pixel-art avatars. A captioned style set: each image ships "
- "a caption, so the adapter learns the pixel look without a trigger word."
- ),
+ "description": "100 captioned pixel-art avatars. A style set, no trigger needed.",
"license": "cc0-1.0",
"image_cap": 100,
"suggested_trigger": None,
diff --git a/studio/frontend/src/features/images/train/diffusion-train-panel.tsx b/studio/frontend/src/features/images/train/diffusion-train-panel.tsx
index 14921462ae..5dbc545a7c 100644
--- a/studio/frontend/src/features/images/train/diffusion-train-panel.tsx
+++ b/studio/frontend/src/features/images/train/diffusion-train-panel.tsx
@@ -77,7 +77,7 @@ const FAMILY_PRESETS: FamilyPreset[] = [
label: "FLUX.1-dev (12B)",
base_repos: ["black-forest-labs/FLUX.1-dev"],
defaults: { rank: 16, lr: 0.0001, resolution: 512 },
- vram_note: "Needs a free Hugging Face license: accept it on the model page, then add your token.",
+ vram_note: "Gated: needs its license and your HF token.",
gated: true,
},
{
@@ -85,21 +85,21 @@ const FAMILY_PRESETS: FamilyPreset[] = [
label: "Qwen-Image (20B)",
base_repos: ["unsloth/Qwen-Image-2512-unsloth-bnb-4bit", "Qwen/Qwen-Image"],
defaults: { rank: 16, lr: 0.00005, resolution: 512 },
- vram_note: "The biggest model: needs a large GPU. Start at 512px.",
+ vram_note: "The biggest: needs a large GPU. Start at 512px.",
},
{
name: "z-image",
label: "Z-Image-Turbo (6B)",
base_repos: ["unsloth/Z-Image-Turbo-unsloth-bnb-4bit", "Tongyi-MAI/Z-Image-Turbo"],
defaults: { rank: 16, lr: 0.0001, resolution: 768 },
- vram_note: "The smallest and fastest to train. A great first pick.",
+ vram_note: "The smallest and fastest. A good first pick.",
},
{
name: "sdxl",
label: "SDXL (U-Net)",
base_repos: ["stabilityai/stable-diffusion-xl-base-1.0", "stabilityai/sdxl-turbo"],
defaults: { rank: 16, lr: 0.0001, resolution: 1024 },
- vram_note: "The classic. Trains comfortably at 1024px.",
+ vram_note: "The classic. Fine at 1024px.",
},
];
@@ -867,11 +867,11 @@ export function DiffusionTrainPanel({
const precisionLabel = (
m: "nf4" | "bf16" | "int8" | "fp8" | "mxfp8" | "auto",
): string => {
- if (m === "auto") return "Auto (recommended)";
- if (m === "nf4") return "nf4 (4-bit QLoRA, lowest VRAM)";
- if (m === "bf16") return "bf16 (fastest, most VRAM)";
- if (m === "int8") return "int8 (8-bit)";
- if (m === "mxfp8") return "mxfp8 (Blackwell, best at high res/batch)";
+ if (m === "auto") return "Auto";
+ if (m === "nf4") return "nf4 (lowest VRAM)";
+ if (m === "bf16") return "bf16 (fastest)";
+ if (m === "int8") return "int8";
+ if (m === "mxfp8") return "mxfp8 (Blackwell)";
return "fp8 (experimental)";
};
@@ -911,7 +911,7 @@ export function DiffusionTrainPanel({
- How fast the model learns over time. Constant is fine for most runs.
+ Constant is fine for most runs.
{lrScheduler !== "constant" &&
@@ -934,7 +934,7 @@ export function DiffusionTrainPanel({
- Saves a lot of GPU memory in exchange for slightly slower steps.
+ Less VRAM, slightly slower steps.
@@ -971,8 +971,7 @@ export function DiffusionTrainPanel({
<>This GPU cannot train this model family.
) : (
<>
- How the base model is stored while training. Auto picks the best fit
- for your GPU.
+ Auto picks the best fit for your GPU.
{basePrequantized && (
<> This base is already 4-bit, so only nf4/auto apply.
)}
@@ -997,7 +996,7 @@ export function DiffusionTrainPanel({
- How the math runs during training. bf16 is right for modern GPUs.
+ bf16 is right for modern GPUs.
)}
@@ -1020,7 +1019,7 @@ export function DiffusionTrainPanel({
- Warms up once at the start, then every training step runs faster.
+ Slower first step, faster after.
)}
@@ -1038,8 +1037,7 @@ export function DiffusionTrainPanel({
Train a LoRA
- Teach an image model a style, character, or subject from your own images. The
- finished adapter shows up in the Create tab's LoRA picker.
+ Teach a model a style or subject from your own images.
@@ -1081,7 +1079,7 @@ export function DiffusionTrainPanel({
{effectiveBase === CUSTOM_BASE && (
- 10-50 images are plenty. Captions are optional: without them, the trigger
- prompt below describes every image.
+ 10-50 images is plenty. Captions are optional.
) : (
@@ -1219,9 +1216,8 @@ export function DiffusionTrainPanel({
/>
)}
{selectedDataset.caption_count === 0 && !gridOpen && (
-
- No captions yet: the trigger prompt below will describe every image,
- or open Review captions to write your own.
+
+ No captions yet, so the trigger prompt describes every image.
)}
@@ -1239,8 +1235,8 @@ export function DiffusionTrainPanel({
just labels the output) */}
{fullyCaptioned ? (
- All {selectedDataset?.image_count} images have captions - no trigger prompt needed.
- The style applies to any prompt after training.
+ All {selectedDataset?.image_count} images have captions, so no trigger prompt
+ is needed.
) : (
@@ -1252,8 +1248,7 @@ export function DiffusionTrainPanel({
className="h-8 text-xs"
/>
- What every image is described as while training. Put these words in a
- prompt later to get the style back.
+ The words you will use later to get this style back.
- The name this LoRA gets in the Create tab's picker.
+ Its name in the Create tab's picker.
@@ -1370,13 +1365,12 @@ export function DiffusionTrainPanel({
Training settings
- Applied when you press Start training
+ Applied on Start training
{trainingSettings}
- Once training starts, live progress and the Training Loss / Gradient Norm
- charts take over this area.
+ Progress and charts take over here once training starts.