Maxilicious20/EchoVision-1.5-Pro

EchoVision 1.5 Pro

EchoVision 1.5 Pro is a text-to-image latent diffusion model with a denoising U-Net trained from scratch. It generates 512 × 512 images from natural-language prompts. The model uses a pretrained, frozen CLIP text encoder for prompt conditioning and a pretrained VAE for image decoding; the diffusion U-Net itself was not initialized from pretrained diffusion weights.

Model details

Property	Value
Model type	Text-conditioned latent diffusion
Denoising network	U-Net trained from scratch
Denoiser checkpoint tensor elements	139,680,132 (~139.68M, including any buffers)
Final training step	245,040
Checkpoint epoch field	13
Training GPU	NVIDIA GeForce RTX 4060
Training image resolution	512 × 512
Latent shape	4 × 64 × 64
Text conditioning	77 × 512
Text encoder	openai/clip-vit-base-patch32
VAE	stabilityai/sd-vae-ft-mse
Saved U-Net weight dtype	FP32
Framework at checkpoint inspection	PyTorch 2.14.1+cu130

The 139,680,132 figure counts tensor elements in the saved U-Net state dictionary; it may include non-trainable buffers. The pretrained text encoder and VAE are not included in that figure.

Architecture

EchoVision 1.5 Pro generates images in latent space. A frozen CLIP encoder converts the input prompt into text embeddings, while the denoising U-Net uses these embeddings and the diffusion timestep to predict noise in a 4-channel latent. Repeated denoising steps produce a latent image, which is decoded into RGB pixels by the VAE.

The U-Net uses a multiscale encoder-decoder architecture with residual processing, skip connections, and text cross-attention. The training objective is diffusion noise prediction (epsilon prediction).

Training

The training cache contains 126,099 image-caption-derived entries at 512 × 512 resolution, with 4 × 64 × 64 image latents and 77 × 512 text embeddings. The denoising U-Net was trained from scratch on this dataset.

The final checkpoint records 245,040 optimizer steps and an epoch field of 13. Training was performed on an NVIDIA GeForce RTX 4060.

The development training configuration used:

Setting	Value
Batch size	8
Optimizer	AdamW
Learning rate	1e-4 initial setting, with warmup and cosine schedule
Diffusion timesteps	1,000
Noise schedule	Scaled-linear
Beta start / end	0.0001 / 0.02
Prediction target	Epsilon (noise)
Conditioning dropout	10%
Latent scaling factor	0.18215

These settings describe the development training configuration; the checkpoint metadata report did not expose the final embedded training settings.

Intended use

EchoVision 1.5 Pro is intended for experimental text-to-image generation, creative image synthesis, and research into latent diffusion models trained from scratch.

Limitations

Output quality and prompt adherence vary with subject matter and prompt complexity. Detailed geometry, multiple interacting objects, anatomy, and text within images may be imperfect. The model may reflect biases present in its training data. No standardized quantitative benchmark results are reported for this release.

License

Apache License 2.0 applies to the released EchoVision model files. The pretrained VAE and CLIP components are separate upstream models governed by their own terms.