Update README.md

Files changed (1) hide show

README.md CHANGED Viewed

@@ -53,7 +53,7 @@ Trained via Ostris' [ai-toolkit](https://replicate.com/ostris/flux-dev-lora-trai
 For this training experiment, we first spent many days rigorously translating the textual elements (slogans, captions, titles, inset poems, speech fragments, etc), with form/signification/rhymes intact, throughout every image subsequently used for training.  <br>
 These translated textographic elements were, furthermore, re-placed by us into their original visual contexts, using fonts matched up to the sources.  <br>
 We then manually composed highly detailed paragraph-long captions, wherein we detailed both the graphic and the textual content of each piece, its layout, as well as the most intuitive/intended apprehension of each composition. <br>
-This version of the resultent LoRA was trained on our custom Schnell-based checkpoint (Historic Color 2), available [here in fp8 Safetensors](https://huggingface.co/AlekseyCalvin/HistoricColorSoonrFluxV2/tree/main) and [here in Diffusers format](https://huggingface.co/AlekseyCalvin/HistoricColorSoonr_v2_FluxSchnell_Diffusers). <br>
 The training went for 5000 steps at a DiT Learning Rate of .00002, batch 1, with the ademamix8bit optimizer, and both text encoders trained alongside the DiT!<br>
 No synthetic data was used for the training, nor any auto-generated captions! Everything was manually and attentively pre-curated with a deep respect for the sources used.  <br>

 For this training experiment, we first spent many days rigorously translating the textual elements (slogans, captions, titles, inset poems, speech fragments, etc), with form/signification/rhymes intact, throughout every image subsequently used for training.  <br>
 These translated textographic elements were, furthermore, re-placed by us into their original visual contexts, using fonts matched up to the sources.  <br>
 We then manually composed highly detailed paragraph-long captions, wherein we detailed both the graphic and the textual content of each piece, its layout, as well as the most intuitive/intended apprehension of each composition. <br>
+This second version of the resultent LoRA was trained on our custom Schnell-based checkpoint (Historic Color 2), available [here in fp8 Safetensors](https://huggingface.co/AlekseyCalvin/HistoricColorSoonrFluxV2/tree/main) and [here in Diffusers format](https://huggingface.co/AlekseyCalvin/HistoricColorSoonr_v2_FluxSchnell_Diffusers). <br>
 The training went for 5000 steps at a DiT Learning Rate of .00002, batch 1, with the ademamix8bit optimizer, and both text encoders trained alongside the DiT!<br>
 No synthetic data was used for the training, nor any auto-generated captions! Everything was manually and attentively pre-curated with a deep respect for the sources used.  <br>