<h1 id="balanced-clip-(1m)">Balanced CLIP (1M)</h1>Training CLIP-G took &gt;15KwH of energy, CLIP-L took far less &lt;1KwHThe full negative reinforcement (Cosine Dissimilarity) is available on my <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/Felldude/22k_CLIP_Negitive">huggingface</a>, this was paired with a positive reinforcement (Contrastive Loss) using the full frozen vision model in latent space.PONY CLIP-L has a further 10 epochs using ASGD for very fine-tuned loss.

Balanced CLIP (Updated)

balancedCLIPUpdated_ponyCLIPG.safetensors

Files with SHA256: 934907853ec0a444d9b450b7d7e53f2981255eb4b52713350c077aa1db7b7d4f